Statistical Analysis of Management Data

Statistical Analysis of Management Data

This page intentionally left blank

Statistical Analysis of Management Data

Hubert Gatignon The Claude Janssen Chaired Professor of Business Administration and Professor of Marketing INSEAD

KLUWER ACADEMIC PUBLISHERS NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW

eBook ISBN: Print ISBN:

0-306-48165-0 1-4020-7315-1

©2003 Kluwer Academic Publishers New York, Boston, Dordrecht, London, Moscow Print ©2003 Kluwer Academic Publishers Dordrecht All rights reserved No part of this eBook may be reproduced or transmitted in any form or by any means, electronic, mechanical, recording, or otherwise, without written consent from the Publisher Created in the United States of America Visit Kluwer Online at: and Kluwer's eBookstore at:

http://kluweronline.com http://ebooks.kluweronline.com

To my daughters, Aline and Valérie

This page intentionally left blank

Contents

Preface 1 Introduction 1.1 Overview 1.2 Objectives 1.2.1 Develop the Student’s Knowledge of the Technical Details of Various Techniques for Analyzing Data 1.2.2 Expose the Students to Applications and “Hand-on” Use of Various Computer Programs for Carrying Out Statistical Analyses of Data 1.3 Types of Scales 1.3.1 Definition of Different Types of Scales 1.3.2 The Impact of the Type of Scale on Statistical Analysis 1.4 Topics Covered 1.5 Pedagogy References 2 Multivariate Normal Distribution 2.1 Univariate Normal Distribution 2.2 Bivariate Normal Distribution 2.3 Generalization to Multivariate Case 2.4 Tests About Means 2.4.1 Sampling Distribution of Sample Centroids 2.4.2 Significance Test: One-sample Problem 2.4.3 Significance Test: Two-sample Problem 2.4.4 Significance Test: K-sample Problem 2.5 Examples 2.5.1 Test of the Difference Between Two Mean Vectors – One-Sample Problem 2.5.2 Test of the Difference Between Several Mean Vectors – K -sample Problem 2.6 Assignment References Basic Technical Readings Application Readings

xiii 1 1 2 2

2 3 4 4 5 6 8 9 9 9 11 12 12 13 15 17 19 19 21 25 27 27 27

viii

Contents

3 Measurement Theory: Reliability and Factor Analysis 3.1 Notions of Measurement Theory 3.1.1 Definition of a Measure 3.1.2 Parallel Measurements 3.1.3 Reliability 3.1.4 Composite Scales 3.2 Factor Analysis 3.2.1 Axis Rotation 3.2.2 Variance Maximizing Rotations (Eigenvalues/vectors) 3.2.3 Principal Component Analysis 3.2.4 Factor Analysis 3.3 Conclusion – Procedure for Scale Construction 3.3.1 Exploratory Factor Analysis 3.3.2 Confirmatory Factor Analysis 3.3.3 Reliability-Coefficient Application Examples 3.4 Assignment 3.5 References Basic Technical Readings Application Readings

29 29 29 29 29 30 33 33 34 37 38 43 43 44 44 44 53 54 54 54

4 Multiple Regression with a Single Dependent Variable 4.1 Statistical Inference: Least Squares and Maximum Likelihood 4.1.1 The Linear Statistical Model 4.1.2 Point Estimation 4.1.3 Maximum Likelihood Estimation 4.1.4 Properties of Estimator 4.2 Pooling Issues 4.2.1 Linear Restrictions 4.2.2 Pooling Tests and Dummy Variable Models 4.2.3 Strategy for Pooling Tests 4.3 Examples of Linear Model Estimation with SAS 4.4 Assignment References Basic Technical Readings Application Readings

55

5 System of Equations 5.1 Seemingly Unrelated Regression (SUR) 5.1.1 Set of Equations with Contemporaneously Correlated Disturbances 5.1.2 Estimation 5.1.3 Special Cases

79 79

55 55 57 59 61 65 65 67 69 71 75 76 76 77

79 81 82

Contents

5.2 A System of Simultaneous Equations 5.2.1 The Problem 5.2.2 Two Stage Least Squares: 2SLS 5.2.3 Three Stage Least Squares: 3SLS 5.3 Simultaneity and Identification 5.3.1 The Problem 5.3.2 Order and Rank Conditions 5.4 Summary 5.4.1 Structure of Matrix 5.4.2 Structure of Matrix 5.4.3 Test of Covariance Matrix 5.4.4 3SLS versus 2SLS 5.5 Examples Using SAS 5.5.1 Seemingly Unrelated Regression Example 5.5.2 Two Stage Least Squares Example 5.5.3 Three Stage Least Squares Example 5.6 Assignment References Basic Technical Readings Application Readings 6 Categorical Dependent Variables 6.1 Discriminant Analysis 6.1.1 The Discriminant Criterion 6.1.2 Discriminant Function 6.1.3 Classification and Fit 6.2 Quantal Choice Models 6.2.1 The Difficulties of the Standard Regression Model with Categorical Dependent Variables 6.2.2 Transformational Logit 6.2.3 Conditional Logit Model 6.2.4 Fit Measures 6.3 Examples 6.3.1 Example of Discriminant Analysis Using SAS 6.3.2 Example of Multinomial Logit – Case 1 Analysis Using LIMDEP 6.3.3 Example of Multinomial Logit – Case 2 Analysis Using LOGIT.EXE 6.3.4 Example of Multinomial Logit – Case 2 Analysis Using LIMDEP 6.4 Assignment References Basic Technical Readings Application Readings

ix

83 83 87 87 88 88 89 91 91 92 92 93 93 93 99 101 103 104 104 104 105 105 105 108 109 112 112 113 117 120 121 121 128 130 133 135 136 136 136

x

Contents

7 Rank Ordered Data 7.1 Conjoint Analysis – MONANOVA 7.1.1 Effect Coding Versus Dummy Variable Coding 7.1.2 Design Programs 7.1.3 Estimation of Part-worth Coefficients 7.2 Ordered Probit 7.3 Examples 7.3.1 Example of MONANOVA Using PC-MDS 7.3.2 Example of Conjoint Analysis Using SAS 7.3.3 Example of Ordered Probit Analysis Using LIMDEP 7.4 Assignment References Basic Technical Readings Application Readings

137 137 137 142 143 144 147 147 151 151 154 156 156 156

8 Error in Variables – Analysis of Covariance Structure 8.1 The Impact of Imperfect Measures 8.1.1 Effect of Errors-in-variables 8.1.2 Reversed Regression 8.1.3 Case with Multiple Independent Variables 8.2 Analysis of Covariance Structures 8.2.1 Description of Model 8.2.2 Estimation 8.2.3 Model Fit 8.2.4 Test of Significance of Model Parameters 8.2.5 Simultaneous Estimation of Measurement Model Parameters with Structural Relationship Parameters Versus Sequential Estimation 8.2.6 Identification 8.3 Examples 8.3.1 Example of Confirmatory Factor Analysis 8.3.2 Example of Model to Test Discriminant Validity Between Two Constructs 8.3.3 Example of Structural Model with Measurement Models 8.4 Assignment References Basic Technical Readings Application Readings

159 159 159 161 162 163 163 165 168 170

9 Analysis of Similarity and Preference Data 9.1 Proximity Matrices 9.1.1 Metric versus Non-metric Data 9.1.2 Unconditional versus Conditional Data

253 253 253 254

171 171 171 172 178 221 250 251 251 251

Contents

9.2

9.3 9.4 9.5

9.6

9.1.3 Derived Measures of Proximity 9.1.4 Alternative Proximity Matrices Problem Definition 9.2.1 Objective Function 9.2.2 Stress as an Index of Fit 9.2.3 Metric 9.2.4 Minimum Number of Stimuli 9.2.5 Dimensionality 9.2.6 Interpretation of MDS Solution 9.2.7 The KYST Algorithm Individual Differences in Similarity Judgments Analysis of Preference Data 9.4.1 Vector Model of preferences 9.4.2 Ideal Point Model of Preference Examples 9.5.1 Example of KYST 9.5.2 Example of INDSCAL 9.5.3 Example of PROFIT (Property Fitting) Analysis 9.5.4 Example of MDPREF 9.5.5 Example of PREFMAP Assignment References Basic Technical Readings Application Readings

Appendices Appendix A: Rules in Matrix Algebra Vector and Matrix Differentiation Kronecker Products Appendix B: Statistical Tables Cumulative Normal Distribution Chi-Squared Distribution F Distribution Appendix C: Description of Data Sets The MARKSTRAT® Environment Survey Indup Panel Scan Index

xi

254 254 255 256 256 257 257 258 258 259 260 260 261 261 261 262 266 275 285 292 306 307 307 307 309 309 309 309 310 310 311 313 314 315 318 318 324 325 329

This page intentionally left blank

Preface I am very indebted to a number of people without whom I would not have envisioned this book. First, Paul Green helped me tremendously in the preparation of the first doctoral seminar I taught at the Wharton School. The orientations and objectives set for that book reflect those he had for the seminar on data analysis which he used to teach before I did. A second individual, Lee Cooper at UCLA, was determinant in the approach I used for teaching statistics. As my first teacher of multivariate statistics, the exercise of having to program all the methods in APL taught me the benefits of such an approach for the complete understanding of this material. Finally, I owe a debt to all the doctoral students in the various fields of management, both at Wharton and INSEAD, who have, by their questions and feedback, helped me develop this approach. I hope it will benefit future students in learning these statistical tools, which are basic to academic research in the field of management especially. Special thanks go to Bruce Hardie who helped me put together some of the data bases and to Frédéric Dalsace who carefully identified sections that needed further explanation and editing. Also, my research assistant at INSEAD, Gueram Sargsyan was instrumental in preparing the examples used in this manual to illustrate the various methods.

This page intentionally left blank

1. Introduction

1.1 Overview This book covers multivariate statistical analyses that are important for researchers in all fields of management whether finance, production, accounting, marketing, strategy, technology or human resources management. Although multivariate statistical techniques such as those described in this book play key roles in fundamental disciplines of the social sciences (e.g., economics and econometrics or psychology and psychometrics), the methodologies particularly relevant and typically used in management research are the center of focus of this study. This book is especially designed to provide doctoral students with a theoretical knowledge of the basic concepts underlying the most important multivariate techniques and with an overview of actual applications in various fields. The book addresses both the underlying mathematics and problems of application. As such, a reasonable level of competence in both statistics and mathematics is needed. This book is not intended as a first introduction to statistics and statistical analysis. Instead, it assumes that the student is familiar with basic statistical techniques. The book presents the techniques in a fundamental way but in a format accessible to students in a doctoral program, to practicing academicians and data analysts. With this in mind, it may be recommended to review some basic statistics and matrix algebra such as provided in the following: Green, Paul E. (1978), Mathematical Tools for Applied Multivariate Analysis, New York, NY: Academic Press [Chapters 2 to 4]. Maddala, G. S. (1977), Econometrics, New York, NY: McGraw Hill, Inc. [Appendix A]. This book offers a clear, succinct exposition of each technique with emphasis on when each technique is appropriate and how to use it. The focus is on the essential aspects that a working researcher will encounter. In short, the focus is on using multivariate analysis appropriately through understanding of the foundations of the methods to gain valid and fruitful insights into management problems. This book presents methodologies for analyzing primary or secondary data typically used by academics as well as analysts in management research and provides an opportunity for the researcher to have hands-on experience with such methods.

2

Statistical Analysis of Management Data

1.2

Objectives

The main objectives of this book are to: 1. Develop the student’s knowledge of the technical details of various techniques for analyzing data. 2. Expose students to applications and “hands-on” use of various computer programs. This experience will make it possible for students to carry out statistical analyses of their own data. Commonly used software is used throughout the book as much as possible across methodologies to avoid having to learn multiple systems with their own, specific data manipulations and instructions. However, not a single data analysis software performs all the analyses presented in the book. Therefore, three basic statistical packages are used: SAS, LIMDEP and LISREL.

1.2.1

Develop the Student’s Knowledge of the Technical Details of Various Techniques for Analyzing Data

The first objective is to prepare the researcher with the basic technical knowledge required for understanding the methods, as well as their limitations. This requires a thorough understanding of the fundamental properties of the techniques. Basic knowledge means that the book will not go into the advanced issues of the methodologies. This should be acquired through specialized, more advanced books on the specific topics. Nevertheless, this book should provide enough detail for what is the minimum knowledge expected from a doctoral candidate in management studies. “Basic” should not be interpreted as a lower level of technical expertise. It is used to express the minimum knowledge expected from an academic researcher in management. The objective is to train the reader to understand the technique, to be able to use it and to have the sufficient knowledge to understand more advanced material about the technique that can be found in other books afterwards.

1.2.2

Expose the Students to Applications and “Hand-on” Use of Various Computer Programs for Carrying Out Statistical Analyses of Data

While the basic statistical theories corresponding to the various types of analysis are necessary, they are not sufficient to do research. The use of any technique requires the knowledge of the statistical software corresponding to these analyses. It is indispensable that students learn both the theory and the practice of using these methods at the same time. A very effective, albeit time consuming way to make sure that the intricacies of a technique are

Introduction

3

mastered is by programming the software oneself. A quicker way is to make sure that the use of the software coincides with the learning of the theory by associating application examples with the theory and by doing some analysis oneself. This is why, in this book, each chapter is made of four parts. The first part of any chapter presents the methods from a theoretical point of view with the various properties of the method. The second part shows an example of an analysis with instructions on how to use a particular software program appropriate for that analysis. The third part gives an assignment so that students can actually practice the method of analysis. The data sets for these assignments are described in Appendix C and can be downloaded from the web at: http://www.insead.edu/~gatignon. Finally, the fourth part consists of references of articles which use such techniques appropriately, and which serve as templates. Selected readings could have been reprinted in this book for each application. However, few articles illustrate all the facets of the techniques. By providing a list of articles, each student can choose the applications corresponding best to his or her interests. By accessing multiple articles in the area of interest, the learning becomes richer. All these articles illustrating the particular multivariate techniques used in empirical analysis are drawn from the major research journals in the field of management.

1.3 Types of Scales

Data used in management research are obtained from existing sources (secondary data) such as data published by Ward for automobile sales in the USA or from vendors who collect data such as panel data. Data are also collected for the explicit purpose of the study (primary data): survey data, scanner data, panels. In addition to this variety of data sources, differences in the type of data which are collected can be critical for their analysis. Some data are continuous measures as, for example, the age of a person, with an absolute starting point at birth or the distance between two points. Some commonly used data do not have such an absolute starting point. Temperature is an example of such a measure. Yet in both cases, i.e., temperatures and distances, multiple units of measurement exist throughout the world. These differences are critical because the appropriateness of data analysis methods varies depending on the type of data at hand. In fact, very often the data may have to be collected in a certain way in order to be able to test hypotheses using the appropriate methodology. Failure to collect the appropriate type of data should prevent performing the test. In this chapter, we discuss the different types of scales which can be found in measuring variables used in management research.

4

Statistical Analysis of Management Data

1.3.1 Definition of Different Types of Scales

Scales are quantitative measures of a particular construct, usually not observed directly. Four basic types of scales can categorize management measurements: Ratio Interval Rank order or ordinal Categorical or nominal 1.3.2

The Impact of the Type of Scale on Statistical Analysis

The nature of analysis depends in particular on the scale of the variable(s). Table 1.1 summarizes the most frequently used statistics which are permissible

Introduction

5

according to the scale type. The order of the scales in the table from Nominal to Ratio is hierarchical in the sense that statistics which are permissible for a scale above are also permissible for the scale in question. For example, a median is a legitimate statistic for an ordinal scale variable but is also legitimate for an interval or ratio scale. The reverse is not true; for example, a mean is not legitimate for an ordinal scale.

1.4

Topics Covered

The methods presented in this book cover the major methods of analysis which have been used in the recent management research literature. A survey of the major journals in the various fields of management was done to identify these methods. This analysis revealed interesting observations. It is striking that the majority of the analyses involve the estimation of a single equation or of several equations independently of one another. Analyses involving a system of equations represent a very small percentage of the analyses performed in these articles. This appears at first hand surprising given the complexity of management phenomena. Possibly some of the simultaneous relationships analyzed are reflected in methodologies which consider explicitly measurement errors; these techniques appear to have grown over the recent years. Factor analysis is still an important analysis found in a significant proportion of the studies, typically to verify the unidimensionality of the constructs measured. Choice modeling has been an important topic, especially in Marketing but also in the other fields of Management, with studies estimating probit or logit models. A still very small percentage of articles use these models for ordered choice data (i.e., where the data reflects only the order in which brands are preferred from best to worse). Analysis of proximity data concerns few studies. Therefore, the following topics were selected. They have been classified according to the type of the key variable or variables which is or are the center of the interest in the analysis. Indeed, as discussed in Chapter 2, the nature of the criterion (also called dependent or endogenous) variable(s) determines the type of statistical analysis which may be performed. Consequently, the first issue to be discussed concerns the nature and properties of variables and the process of generating scales with the appropriate statistical procedures. Then, follow the various statistical methods of data analysis. Introduction to multivariate statistics and tests about means Multivariate Analysis of Variance Multiple item measures Reliability

6

Statistical Analysis of Management Data

Factor Analysis Principle Component Analysis Exploratory Factor Analysis Confirmatory Factor Analysis Single equation econometrics Ordinary Least Squares Generalized Least Squares Pooling Tests System of equations econometrics Seemingly Unrelated Regression Two Stage Least Squares Three Stage Least Squares Categorical dependent variables Discriminant Analysis Quantal choice Models: Logit Rank ordered data Conjoint Analysis Ordered Probit Analysis of covariance structure LISREL Analysis of similarity data Multidimensional Scaling

1.5 Pedagogy

There are three key learning experiences necessary to be able to achieve these objectives: 1. the knowledge of sufficient statistical theory to be able to understand the methodologies, when they are applicable, and when they are not appropriate.

Introduction

7

2. the ability to perform such analyses with the proper statistical software, 3. the understanding of how these methodologies have been applied in management research.

This book differs from others in that no other book on multivariate statistics or data analysis addresses the specific needs of doctoral education. The three aspects mentioned above are weighted differently. This book emphasizes the first aspect of the methodology itself by providing the mathematical and statistical analyses necessary to fully understand them. This can be contrasted with other books that prefer primarily or exclusively a verbal description of the method. This book favors the understanding of the rationale for modeling choices, issues and problems. While the verbal description of a method may be better accessible to a wider audience, it is often more difficult to follow the rationale, which is based on mathematics. For example, it is difficult to understand the problem of multicollinearity without understanding the effect on the determinant of the covariance matrix which needs to be inverted. The learning that results from verbal presentation tends, therefore, to be more mechanical. This book also differs in that, instead of choosing a few articles to illustrate the applications of the methods, as would be found in a book of readings (sometimes with short introductions), a list of application articles is provided from which the reader can choose. Articles tend to be relatively easy to access, especially with services available through the WEB. The list of references covers a large cross section of examples and a history of the literature in this domain. Finally, the examples of analyses are relatively self explanatory and, although some explanations of the statistical software used are provided with each example, this book does not intend to replace the instruction manuals of those particular software packages. The reader is referred to those for details. In summary, this book puts the accent on the first aspect of the understanding of the statistical methodology while providing enough information for the reader to develop skills in performing the analyses and in understanding how to apply them to management research problems. More specifically, the learning of this material involves two parts: the learning of the statistical theory behind the technique and the learning of how to use the technique. Although there may be different ways to combine these two experiences, it is recommended to first learn the theory by reading the sections where the methodologies are presented and discussed. Then, the statistical computer package (e.g., SAS, LIMDEP, LISREL, and other specialized packages) used to apply the methodology is presented in the context of an example. Students can then apply the technique using the data sets available from http://www.insead.edu/~gatignon. Finally, application issues can be illustrated by other applications found in prior research and listed at the end of each chapter.

8

Statistical Analysis of Management Data

In addition to the books and articles included with each chapter, the following books are highly recommended to further develop someone’s skills in different methods of data analysis. Each of these books is more specialized and covers only a subset of the methods presented in this book. However, they are indispensable complements to become proficient in the techniques used in research. References Greene, W. H. (1993), Econometric Analysis, New York: MacMillan Publishing Company. Hanssens, D. M., L. J. Parsons and R. L. Shultz (1990), Market Response Models: Econometric and Time Series Analysis, Norwell, MA: Kluwer Academic Publishers. Judge, G. G., W. E. Griffiths, R. C. Hill, H. Lutkepohl and T.-C. Lee (1985), The Theory and Practice of Econometrics, New York, NY: John Wiley & Sons.

2. Multivariate Normal Distribution In this chapter, we will define the univariate and multivariate normal distribution density functions and then we will discuss the tests of differences of means for multiple variables simultaneously across groups. 2.1 Univariate Normal Distribution Just to refresh memory, in the case of a single random variable, the probability distribution or density function of that variable is represented by Equation (2.1):

2.2

Bivariate Normal Distribution

The bivariate distribution represents the joint distribution of two random variables. The two random variables and are related to each other in the sense that they are not independent of each other. This dependence is reflected by the correlation between the two variables and The density function for the two variables jointly is:

This function can be represented graphically as in Figure 2.1. The Isodensity contour is defined as the set of points for which the values of and give the same value for the density function This contour is given by Equation (2.3) for a fixed value of C, which defines a constant probability:

Equation (2.3) defines an ellipse with centroid This ellipse is the locus of points representing the combinations of the values of and with the same probability, as defined by the constant C (Figure 2.2).

10

Statistical Analysis of Management Data

For various values of C, we get a family of concentric ellipses (at a different cut, i.e., cross section of the density surface with planes at various elevations) (see Figure 2.3). The angle depends only on the values of and but is independent of C. The higher the correlation between and the steeper the line going through the origin with angle i.e., the bigger the angle.

Multivariate Normal Distribution

11

2.3 Generalization to Multivariate Case Let us represent the bivariate distribution in matrix algebra notation in order to derive the generalized format for more than two random variables. The covariance matrix of can be written as:

The determinant of the matrix

is:

Equation (2.3) can now be re-written as:

where

Note that

× matrix of cofactors.

12

Statistical Analysis of Management Data

Let

then chi-square variate. Also, because consequently,

which is a quadratic form equation and is, therefore, a and

The bivariate distribution function can be now expressed in matrix notation as: Now, more generally with p random variables

let

The density function is:

For a fixed value of the density an ellipsoid is described. Let The inequality defines any point within the ellipsoid. 2.4 Tests About Means 2.4.1 2.4.1.1

Sampling Distribution of Sample Centroids Univariate distribution

A random variable is normally distributed with mean

and variance

After n independent draws, the mean is randomly distributed with mean and variance

Multivariate Normal Distribution 2.4.1.2

13

Multivariate distribution

In the multivariate case with p random variables where x = x is normally distributed following the multivariate normal distribution with mean and covariance :

The mean vector for the sample of size n is denoted:

This sample mean vector is normally distributed with a multivariate normal distribution with mean and covariance

2.4.2 2.4.2.1

Significance Test: One-sample Problem Univariate test

The univariate test is illustrated in the following example. Let us test the hypothesis that the mean is 150 (i.e., with the following information: Then, the z score can be computed:

At (95% confidence interval), z = 1.96, as obtained from a normal distribution table. Therefore, the hypothesis is rejected. The confidence interval is

This interval excludes 150. The hypothesis that is rejected. If the variance had been unknown, the t statistic would have been used:

where s is the observed sample standard deviation.

Statistical Analysis of Management Data

14

2.4.2.2

Multivariate test with known

Let us take an example with two random variables:

The hypothesis is now about the mean values stated in terms of the two variables jointly:

At the alpha level of 0.05, the value of the density function can be written as below, which follows a chi-squared distribution at the specified significance level

Computing the value of the statistics,

The critical value at an alpha value of 0.05 with 2 degrees of freedom is provided by tables:

The observed value is greater than the critical value. Therefore, the hypothesis that is rejected. 2.4.2.3

Multivariate test with unknown

Just as in the univariate case, is replaced with the sample value S/ (n – 1), where S is the sums-of-squares-and-cross-products (SSCP) matrix, which provides an unbiased estimate of the covariance matrix. The following statistics

Multivariate Normal Distribution

15

are then used to test the hypothesis:

where, if

Hotelling showed that

Replacing

by its expression given above:

Consequently, the test is performed by computing the expression above and comparing its value with the critical value obtained in an F table with p and n – p degrees of freedom. 2.4.3 2.4.3.1

Significance Test: Two-sample Problem Univariate test

Let us define and the means of a variable on two unrelated samples. The test for the significance of the difference between the two means is given by

where

is the pooled within groups variance. It is an estimate of the assumed common variance of the two populations.

16

Statistical Analysis of Management Data

2.4.3.2

Let

Multivariate test

be the mean vector in sample 1 =

and similarly for

sample 2. We need to test the significance of the difference between and . We will consider first the case where the covariance matrix, which is assumed to be the same in the two samples, is known. Then we will consider the case where an estimate of the covariance matrix needs to be used. is known (the same in the two samples) In this case, the difference between the two group means is normally distributed with a multivariate normal distribution:

The computations for testing the significance of the differences are similar to those in section 2.4.2.2. using the chi-squared test. is unknown If the covariance matrix is not known, it is estimated using the covariance matrices within each group but pooled. Let W be the within-groups SSCP (sum of squares cross products) matrix. This matrix is computed from the matrix of deviations from the means on all p variables for each of observations (individuals). For each group k,

For each of the 2 groups (each k), the SSCP matrix can be derived:

The pooled SSCP matrix for the more general case of K groups is simply:

In the case of two groups, K is simply equal to 2.

Multivariate Normal Distribution

17

Then, we can apply Hotelling’s T, just as in section 2.4.2.3, where the proper degrees of freedom depending on the number of observations in each group are applied.

2.4.4

Significance Test: K-sample Problem

As in the case of two samples, the null hypothesis is that the mean vectors across the K groups are the same and the alternative hypothesis is that they are different. Let us define Wilk’s likelihood-ratio criterion:

where T = total SSCP matrix, W = within-groups SSCP matrix. W is defined as in Equation (2.25). The total SSCP matrix is the sum of squared cross products applied to the deviations from the grand means (i.e., the overall mean across the total sample with the observations of all the groups for each variable). Therefore, let the mean centered data for group k be noted as:

where is the overall mean of the j’s variate. Bringing the centered data for all the groups in the same data matrix leads to:

18

Statistical Analysis of Management Data

The total SSCP matrix T is then defined as:

Intuitively, if we reduce the space to a single variate so that we are only dealing with variances and no covariances, Wilk’s lambda is the ratio of the pooled within variance to the total variance. If the group means are the same, the variances are equal and the ratio equals one. As the group means differ, the total variance becomes larger than the pooled within group variance. Consequently, the ratio lambda becomes smaller. Because of the existence of more than one variate, which implies more than one variance and covariances, the within SSCP and Total SSCP matrices need to be reduced to a scalar in order to derive a scalar ratio. This is the role of the determinants. However, the interpretation remains the same as for the univariate case. Based on Wilk’s lambda, we will present two statistical tests: Bartlett’s V and Rao’s R. Let n = total sample size across samples, p = number of variables, K = number of groups (number of samples). Bartlett’s V is approximately distributed as a chi-square when n – 1 – (p + K)/2 is large:

Bartlett’s V is relatively easy to calculate and can be used when n – 1 – (p + K)/2 is large. Another test can be applied, as Rao’s R is distributed approximately as an F variate. It is calculated as follows:

where

Multivariate Normal Distribution

2.5 2.5.1

19

Examples

Test of the Difference Between Two Mean Vectors – One-Sample Problem

In this example, the file “MKT_DATA” contains data about the market share of a brand over seven periods, as well as the percentage of distribution coverage and the price of the brand. These data correspond to one market, Norway. The question is to know whether the market share, distribution coverage and prices are similar or different from the data of that same brand for the rest of Europe, i.e., with values of market share, distribution coverage and price respectively of 0.17, 32.28 and 1.39. The data are shown below in Table 2.1. The SAS file showing the SAS code to compute the necessary statistics is shown below in Figure 2.4. The first lines correspond to the basic SAS instructions to read the data from the file. Here, the data file was saved as a text file from Excel. Consequently, the values in the file corresponding to different data points are separated by commas. This is indicated as the delimiter (“dlm”). Also, the data (first observation) starts on line 2 because the first line is used for the names of the variables (as illustrated in Table 2.1). The variable called period is dropped so that only the three variables needed for the analysis are kept in the SAS working data set. The procedure IML is used to perform matrix algebra computations. This file could easily be used for the analysis of different data bases. Obviously, it would be necessary to adapt some of the instructions, especially the file name and path and the variables. Within the IML subroutine, only three things would need to be changed: (1) the variables used for the analysis, (2) the values for the null hypothesis (m_o) and (3) the critical value of the F statistic with the proper degrees of freedom. The results are printed in the output file shown below in Figure 2.5. The critical F statistic with 3 and 4 degrees of freedom at the 0.05 confidence level is 6.591 while the computed value is 588.7, indicating that the hypothesis of no difference is rejected.

20

Statistical Analysis of Management Data

Multivariate Normal Distribution

2.5.2

21

Test of the Difference Between Several Mean Vectors – K-sample Problem

The next example considers similar data for three different countries (Belgium, France and England) for seven periods, as shown in Table 2.2. The question is to know whether the mean vectors are the same for the three countries or not.

22

Statistical Analysis of Management Data

Multivariate Normal Distribution

23

24

Statistical Analysis of Management Data

The SAS file which derived the computations for the test statistics is shown in Figure 2.6. The results are shown in the SAS output below (Figure 2.7). These results indicate that the Bartlett’s V statistic of 82.54 is larger than the critical chi square with 6 degrees of freedom at the 0.05 confidence level (which is 12.59). Consequently, the hypothesis that the mean vectors are the same is rejected. The same conclusion could be derived from the Rao’s R statistic with its value of 55.10, which is larger than the corresponding F value with 6 and 32 degrees of freedom which is 2.399.

Multivariate Normal Distribution

2.6

25

Assignment

In order to practice with these analyses, you will need to use the data bases INDUP and PANEL described in Appendix C. These data bases provide market share and marketing mix variables for a number of brands competing in five market segments. You can test the following hypotheses:

1. The market behavioral responses of a given brand (e.g., awareness, perceptions or purchase intentions) are different across segments, 2. The marketing strategy (i.e., the values of the marketing mix variables) of selected brands is different (perhaps corresponding to different strategic groups). Figure 2.8 shows how to read the data within a SAS file and how to create new files with a subset of the data saved in a format which can be read easily using the examples provided throughout this chapter. Use the model described in the examples above and adapt them to the data base to perform these tests.

26

Statistical Analysis of Management Data

Multivariate Normal Distribution

27

References Basic Technical Readings Tatsuoka, M. M. (1971), Multivariate Analysis: Techniques for Educational and Psychological Research, New York, NY: John Wiley & Sons, Inc.

Application Readings Cool, K. and I. Dierickx (1993), “Rivalry, Strategic Groups and Firm Profitability,” Strategic Management Journal, 14, 47–59. Long, R. G., W. P. Bowers, T. Barnett, et al. (1998), “Research Productivity of Graduates in Management: Effects of Academic Origin and Academic Affiliation,” Academy of Management Journal, 41, 6, 704–714.

This page intentionally left blank

3. Measurement Theory: Reliability and Factor Analysis

In this chapter, we will discuss the issues involved in building measures or scales. We focus the chapter on two types of analysis: (1) the measurement of reliability with Cronbach’s alpha and (2) the verification of unidimensionality using factor analysis. In this chapter, we concentrate on Exploratory Factor Analysis and we only introduce the notion of Confirmatory Factor Analysis. Although an important step in the construction and the evaluation of scales, the analysis required is a special case of the Analysis of Covariance Structures presented in Chapter 8. Consequently, we postpone the estimation of the parameters of Confirmatory Factor Analysis, as well as the examination of convergent and discriminant validity issues to this chapter.

3.1

Notions of Measurement Theory

3.1.1 Definition of a Measure

If T is the true score of a construct and e represents the error associated to the measurement, the measure X is expressed as:

3.1.2

Parallel Measurements

Measures

3.1.3

and

are parallel if they meet the following characteristics:

Reliability

The reliability of a measure is the squared correlation between the measure and the true score: also noted It is also the ratio of the true

30

Statistical Analysis of Management Data

score variance to the measure variance:

This can be demonstrated as follows:

This last equality can be shown as follows:

but which is the numerator of the reliability expression. Let us now express the correlation between the true score and the measure:

Therefore, the reliability can be expressed as the proportion of the observed score variance that is true score variance. The problem with the definition and formulae above is that the variance of the true score is not known since the true score is not observed. This explains the necessity to use multiple measures and to form scales. 3.1.4

Composite Scales

A composite scale is built from using multiple items or components measuring the constructs. This can be represented graphically as in Figure 3.1. Note that by convention, circles represent unobserved constructs and squares identify observable variables or measures.

Measurement Theory: Reliability and Factor Analysis

3.1.4.1

31

Reliability of a two-component scale

In this section, we show that the reliability of a composite scale has a lower bound. This lower bound is coefficient alpha. The two components of the scale are:

The composite scale corresponds to a formative index:

Although, a priori, and appear as different true scores, we will see that they must be positively correlated and we will show the impact of that correlation on the reliability of the scale. As a consequence, it is best to think of these scores as corresponding to different items of a single construct. Computation of coefficient From Equation (3.16), the composite scale defined as:

However, because (equality if parallel test), then it follows that:

This last inequality results from developing the inequality above it:

Since, given a positive correlation between

and

32

Statistical Analysis of Management Data

It follows that:

The left hand side of the inequality above being positive, a fortiori, the right hand side is also positive. It should be noted that this property is only interesting for cases where the items (components) are positively correlated. Indeed, in the case of a negative correlation, the inequality is dominated by the fact that the left hand side is greater or equal to zero. Therefore, in cases of positively correlated items:

Consequently, the reliability has a lower bound which is given by

But

Therefore,

But since

Then and therefore

This demonstrates that there is a lower bound to the reliability. If this lower bound is high enough, this means that the actual reliability is even higher and therefore, the scale is reliable. It is also clear from Equation (3.28) that as the (positive) correlation between the two items or components increases, the portion that is substracted from one decreases so that coefficient alpha increases. If the correlation is zero, then coefficient alpha is zero.

Measurement Theory: Reliability and Factor Analysis 3.1.4.2

33

Generalization to composite measurement with K components

For a scale formed from K components or items:

The reliability coefficient alpha is a generalized form of the calculation above:

is a lower bound estimate of the reliability of the composite scale X, that is of

3.2

Factor Analysis

Factor analysis can be viewed as a method to discover or confirm the structure of a covariance matrix. However, in the case of exploratory factor analysis, the analysis attempts to discover the underlying unobserved factor structure. In the case of confirmatory factor analysis, a measurement model is specified and tested against the observed covariance matrix. Exploratory factor analysis is a special type of rotation. Consequently, rotations are first reviewed in the general context of space geometry. 3.2.1

Axis Rotation

Let us consider Figure 3.2, which shows a set of orthogonal axes and The vector shows an angle relative to Similarly, the vector forms an angle with The rotation corresponds to a linear transformation of x to y. If x is a p-dimensional vector and V is a square matrix of size p by p (which represents the linear weights applied to vector x), then y, the linear transformation of x,

34

Statistical Analysis of Management Data

is also with dimension p. However, orthogonality conditions must be met so that V cannot be any matrix. Therefore, the rotation can be expressed in the following equations:

so that conditions be met for orthogonal rotation. An example of a rotation in a two-dimensional space is given below:

These weights represented in Equations (3.33) and (3.34) are appropriate for an orthogonal rotation because the constraints of orthogonality expressed in Equation (3.32) are respected. Indeed,

These constraints can be expressed in matrix notations as:

This corresponds to the constraint expressed more generally in Equation (3.32).

3.2.2

Variance Maximizing Rotations (Eigenvalues/vectors)

The advantage of an orthogonal rotation is that it enables to represent the same points in a space using different axes but without affecting the covariance matrix which remains unchanged. The idea is now going to be to find a specific rotation or linear transformation which will maximize the variance of the linear transformations. 3.2.2.1

The objective

The objective is, therefore, to find the linear transformation of a vector which maximizes the variance of the transformed variable (of the linear combination) i.e., to find the weights such that if for one observation, the

Measurement Theory: Reliability and Factor Analysis

35

transformation is and for all N observations

the variance of the transformed variable, which is proportional to

is maximized. In other words, the problem is:

This is equivalent, by replacing y with its expression as a linear combination of X, to:

This can be resolved by maximizing the Lagrangian L: Using the derivative rule

Solving these equations provides the eigenvalues and eigenvectors. First we show how to derive the eigenvalues. Then, we will proceed with the calculation of the eigenvectors. Finding the eigenvalues We need to resolve the following system of equations for v and A trivial solution is v = 0. Pre-multiplying by This implies also that, for a non-trivial solution to exist, must not have an inverse because, if it does, v = 0 and gives a trivial solution.

36

Statistical Analysis of Management Data

Therefore, a first condition for a non-trivial solution to Equation (3.46) to exist is that the determinant is zero because the operation shown in Equation (3.47) cannot then be performed:

Equation (3.48) results in a polynomial in of degree p and therefore which has p roots. Following is an example. Let us assume that the covariance matrix is:

Then,

Resolving this second degree equation gives the two roots:

They are the eigenvalues. Finding the eigenvectors Knowing the eigenvalues, the eigenvectors can now be easily computed. For each eigenvalue, there are p equations with p unknown:

subject to normality, i.e., The p unknowns are then straightforward to estimate. 3.2.2.2

Properties of eigenvalues and eigenvectors

Two properties of eigenvectors and eigenvalues are indispensable in order to understand the implications of this rotation:

It is important to understand the proof of this last property because it shows how the covariance matrix can be reconstituted with the knowledge of eigenvectors and eigenvalues.

Measurement Theory: Reliability and Factor Analysis

37

From the first order derivative of the Lagrangian and putting all eigenvectors together)

Premultiplying each side by

gives:

Furthermore, a third property is that the eigenvalue is the variance of the linearly transformed variable y. From Equation (3.51), premultiplying the left-hand side by one obtains for eigenvalue i and eigenvector i:

However, the left-hand side of Equation (3.57) is the variance of the transformed variable

Therefore, the eigenvalue represents the variance of the new variable formed as a linear combination of the original variables.

3.2.3

Principal Component Analysis

The problem in Principal Component Analysis is just what has been described in the prior section. It consists in finding the linear combination that maximizes the variance of the linear combinations of a set of variables (the first linear combination, then the second given that it should be perpendicular to the first, etc.) and to reconstitute the covariance matrix Therefore, the problem is identical to finding the eigenvalues and eigenvectors of the covariance matrix. In Principal Component Analysis, new variables (y) are constructed as exact linear combinations of the original variables. Furthermore, it is a data reduction method in the sense that the covariance matrix can be approximated with a number of dimensions smaller than p, the number of original variables.

38

Statistical Analysis of Management Data

Indeed, from Equation (3.55):

Let to include the eigenvectors corresponding to the r largest eigenvalues and to include the r largest eigenvalues:

Therefore, it can be seen from Equation (3.63) that replacing the small eigenvalues by zero should not affect the ability to reconstitute the variance covariance matrix should approximate S). Consequently, r data points are needed for each i instead of the original p variables. Two points can be made which distinguish Principal Component Analysis from Factor Analysis:

1. The new variables y are determined exactly by the p x variables. There is no noise introduced which may represent some measurement error, as discussed in the section on measurement theory. Factor Analysis introduces this notion of measurement error. 2. The new unobserved variables y are built by putting together the original p variables. Therefore, y is constructed from the original x variables in an index. As opposed to this constitutive index, in Factor Analysis, the observed x variables reflect from the various unobserved variables or constructs. This last distinction between reflective indicators and constitutive indices is developed in the next section. 3.2.4

Factor Analysis

Now that we have explained the difference between Principle Component Analysis and Factor Analysis, we need to distinguish between two different types of Factor Analysis: Exploratory Factor Analysis and Confirmatory Factor Analysis. The basic difference lies in the fact that in Confirmatory Factor Analysis, a structure is proposed in which the observed, measurable variables reflect only specific unobserved constructs while Exploratory factor Analysis allows all measurable variables to reflect from each factor. These two types of Factor Analysis can easily be distinguished by the differences in their graphical representation. Then we will examine the differences analytically.

Measurement Theory: Reliability and Factor Analysis

39

Exploratory Factor Analysis is graphically represented in Figure 3.3 in an example with two unobserved constructs and five observed variables or measures. The unobserved constructs are represented with circles while the measures are represented by squares. The arrows on the left side coming into the measured variable boxes indicate the random measurement errors. In Confirmatory Factor Analysis, only some measures are reflecting specific, individual unobserved constructs, as shown in Figure 3.4. 3.2.4.1

Exploratory Factor Analysis

Exploratory Factor Analysis can be characterized by the fact that it is data driven, as opposed to Confirmatory Analysis which represents a theory of measurement. The purpose of Exploratory Factor Analysis is, in fact, to find or discover patterns which may help understand the nature of the unobserved variables. Consequently, it is a method which, based on patterns of correlations among variables, inductively brings insights into the underlying factors. Considering Figure 4.3, the weights assigned to each arrow linking each factor to each observed variable indicate the extent to which each variable reflects each factor. This can be shown analytically.

40

Statistical Analysis of Management Data

As discussed above, each observed variable is a function of all the factors underlying the structure. For example, for two observed variables and two factors:

where

The variances are one because they are standardized without imposing additional constraints but which enables the identification. This in a sense simply determines the units of measure of the unobserved construct. Let us consider now the consequences that these equations impose on the structure of the covariance matrix of the observed variables.

Using the property that the factors are orthogonal (uncorrelated with a variance of 1):

These equalities follow from the fact that:

Therefore, the variances in the covariance matrix is composed of two components: commonalities and unique components:

in Equation (3.74) represents the proportion of variance explained by the common factors while represents the unique variance. The commonalities are our center of interest because the error variance or unique variances do not contain information about the data structure. This demonstrates that the noise or measurement error needs to be removed although it only affects the variances (the diagonal of the covariance matrix) but not the covariances.

Measurement Theory: Reliability and Factor Analysis

41

More generally, we can represent the data structure as:

where U = diag{u}. R is the matrix of common variance and covariances and U is the matrix of unique variances. In Exploratory factor Analysis, the objective will be to reduce the dimensionality of the R matrix to understand better the underlying factors driving this structure pattern. Four steps are involved in Exploratory Factor Analysis. We discuss each step in turn and then we derive the factor loadings and the factor scores. Estimating commonalities In this first step, we need to remove the unique component of the variance in order to keep the variance explained by the common factors only. In a typical Exploratory Factory Analysis, R is specified as the squared multiple correlations of each variable with the remainder of the variables in the set (i.e., the percentage of explained variance obtained in regressing variable j on the (p – 1) others). U is the residual variances from these regressions. Extracting initial factors The initial factors are obtained by performing a Principal Component Analysis on R

Determining the number of factors The issue is to find the number of factors r < p which are necessary to represent the covariance structure. Following from the properties of eigenvalues and eigenvectors, Let to include the eigenvectors corresponding to the r largest eigenvalues and to include the r largest eigenvalues:

The problem is about finding r so as to account for most of the matrix R. A typically used method is the Scree test which consists in plotting the eigenvalues in the order of their decreasing size. The rules to apply are then: eliminate values less than 1. The rationale for this rule is that each factor should at least account for at least the variance of a single variable, 2. the elbow rule: stop when the curve forms an elbow as shown in Figure 3.5. 1.

42

Statistical Analysis of Management Data

None of these methods should be used blindly. Especially, the rule of the eigenvalue greater than one is the default option on most statistical analysis softwares, including SAS. Indeed, the interpretation of the factors is an important criterion for making sense out of the covariance structure. Rotation to terminal solution The objective for performing a rotation at this stage, using only the retained factors, is to find more easily interpretable factors through rotation. The most commonly used method is the VARIMAX rotation method. With that method, the rotation searches to give the maximum variance of the squared loadings for each factor (in order to avoid problems due to negative loadings). This results in obtaining extreme loadings (very high or very low). Component loadings (for Principle Component Analysis)

because Component scores (for Principle Component Analysis)

3.2.4.2

Confirmatory Factor Analysis

In Confirmatory Factor Analysis, a measurement model is assumed. The objective is to test if the data fit the measurement model. This is, therefore, an ultimate test of the fit of the measurement model to the data.

Measurement Theory: Reliability and Factor Analysis

43

The figure shown in Figure 3.4 can be expressed by the system of equations.

The theoretical covariance matrix of X is given by

If the observed covariance matrix estimated from the sample is S, we need to find the values of the lambdas which will reproduce a covariance matrix as similar as possible to the observed one. Maximum Likelihood estimation is used to minimize How many parameters need to be estimated? In the example shown in Figure 4.4, ten parameters must be estimated: The methodology for estimating these parameters is presented in Chapter 8. 3.3

Conclusion - Procedure for Scale Construction

Scale construction involves several steps. Of those steps, this chapter discussed the following statistical analyses which provide a guide in scale construction. 3.3.1

Exploratory Factor Analysis

Exploratory Factor Analysis can be performed separately for each hypothesized factor. This demonstrates the unidimensionality of each factor. One

44

Statistical Analysis of Management Data

global Factor Analysis can also be performed in order to assess the degree of independence between the factors. 3.3.2

Confirmatory Factor Analysis

Confirmatory factor analysis can be used to assess the overall fit of the entire measurement model and to obtain the final estimates of the measurement model parameters. Although sometimes performed on the same sample as the exploratory factor analysis, when it is possible to collect more data, it is preferable to perform the confirmatory factor analysis on a new sample. 3.3.3

Reliability-Coefficient

In cases where composite scales are developed, this measure is useful to assess the reliability of the scales. Reliabilities of less than 0.7 for academic research and 0.9 for market research are typically not sufficient to warrant further analyses using these composite scales. 3.4

Application Examples

Figure 3.6 illustrates how to compute means and the correlation matrix for a list of variables in SAS. The output is shown on Figure 3.7. A factor analysis on

Measurement Theory: Reliability and Factor Analysis

45

46

Statistical Analysis of Management Data

the same list of variables is requested in Figure 3.8 using the SAS procedure “Factor.” The results are shown in Figure 3.9. This factor analysis of the perception of innovations on nine characteristics are summarized by two factors with eigenvalues greater than one (the default option in SAS) and that explain 89.9% of the variance. The rotated factor patterm shows that Factor 1 groups variables IT1, IT3, IT4, IT6 and IT7, while Factor 2 reflects variables IT5, IT8 and IT9. Variable IT2 does not discriminate well between the two factors, as it loads simultaneously on both, although it loads slightly more on Factor 2. The reliability of the scales (corresponding to the two

Measurement Theory: Reliability and Factor Analysis

47

48

Statistical Analysis of Management Data

Measurement Theory: Reliability and Factor Analysis

49

50

Statistical Analysis of Management Data

Measurement Theory: Reliability and Factor Analysis

51

52

Statistical Analysis of Management Data

Measurement Theory: Reliability and Factor Analysis

53

factors) are then calculated in Figure 3.10 when the variables are first standardized. Those variables with negative loadings are reversed so that each component has the same direction (positive correlations). The results are listed in Figure 3.11 which shows the reliability coefficient alpha for each scale and the improvements that could be obtained by deleting any single variable one at a time. Finally, Figure 3.12 shows how to create a scale composed of these standardized variables, scales that are used in a single analysis of variance example. The corresponding output in Figure 3.13 shows for example the means of the two scales (labeled Tech and MKT) for two levels of the variable RAD. 3.5

Assignment

The assignment consists in developing a composite scale, demonstrating its unidimensionality and computing its reliability. For that purpose, survey data are provided in the file SURVEY.ASC. These data concern items about psychographic variables which contain opinion, attitude and life style characteristics of individuals. The detailed description of the data is given in Appendix C. This type of data is useful for advertising and segmentation purposes. In order to develop a scale, it may be useful to summarize the data using exploratory factor analysis on a wide range of variables. It is important, however, to make sure that only variables which possess the properties necessary

54

Statistical Analysis of Management Data

for the analysis are included. For example, because factor analysis is based on correlations, categorical or ordinal scale variables should be excluded from the analysis, since correlations are not permissible statistics with such scales. The factors need to be interpreted and you can concentrate on a subset of these factors to derive a single or multiple composite scales. An alternative would be to reflect on the questions which seem related and focus on those to develop a scale. This is in essence a mental factor analysis. You need to demonstrate that each of the scales developed are unidimensional (through factor analysis) and that their reliability is sufficiently high. Figure 3.14 lists the SAS file which can be used to read the data. References Basic Technical Readings Bollen, K. and R. Lennox (1991), “Conventional Wisdom on Measurement: A Structural Equation Perspective,” Psychological Bulletin, 110, 2, 305–314. Cortina, J. M. (1993), “What is Coefficient Alpha? An Examination of Theory and Applications,” Journal of Applied Psychology, 78, 1, 98–104. Diamanopoulos, A. and H. M. Winklhofer (2001), “Index Construction with Formative Indicators: An Alternative to Scale Development,” Journal of Marketing Research, 38, 2 (May), 269–277. Green, P. E. (1978), Mathematical Tools for Applied Multivariate Analysis, New York, NY: Academic Press, [Chapter 5 and Chapter 6, section 6.4]. Lord, F. M. and M. R. Novick (1968), Statistical Theories of Mental Test Scores, Reading, MS: AddisonWesley Publishing Company, Inc., [Chapter 4]. Nunnally, J. C. and I. H. Bernstein (1994), Psychometric Theory, Third Edition, New York: McGraw Hill.

Application Readings Aaker, J. L. (1997), “Dimensions of Brand Personality”, Journal of Marketing Research, 34, 3 (August), 347–356. Anderson, E. (1985), “The Salesperson as Outside Agent or Employee: A Transaction Cost Analysis,” Marketing Science, 4 (Summer), 234–254. Anderson, R. and J. Engledow (1977), “A Factor Analytic Comparison of U.S. and German Information Seeker,” Journal of Consumer Research, 3,4, 185–196. Blackman, A. W. (1973), “An Innovation Index Based on Factor Analysis,” Technological Forecasting and Social Change, 4, 301–316. Churchill, G. A., Jr. (1979), “A Paradigm for Developing Better Measures of Marketing Constructs,” Journal of Marketing Research, 16 (February), 64–73. Deshpande, R. (1982), “The Organizational Context of Market Research Use,” Journal of Marketing, 46, 4 (Fall), 91–101. Finn, A. and U. Kayandé (1997), “Reliability Assessment and Optimization of Marketing Measurement,” Journal of Marketing Research, 34, 2 (May), 262–275. Gilbert, F. W. and W. E. Warren (1995), “Psychographic Constructs and Demographic Segments,” Psychology & Marketing, 12, 3 (May), 223–237. Green, S.G., M. B.Gavin and L. Aiman-Smith(1995), “Assessing a Multidimensional Measure of Radical Technological Innovation”, IEEE Transactions on Engineering Management, 42, 3, 203–214. Murtha, T. P., S. A. Lenway and R. P. Bagozzi (1998), “Global Mind-sets and Cognitive Shift in a Complex Multinational Corporation,” Strategic Management Journal, 19, 97–114. Perreault, W. D., Jr. and L. E. Leigh (1989), “Reliability of Nominal Data Based on Qualitative Judgments,” Journal of Marketing Research, 26 (May), 135–148. Zaichowsky, J. L. (1985), “Measuring the Involvement Construct,” Journal of Consumer Research, 12 (December), 341–352.

4. Multiple Regression with a Single Dependent Variable

In this chapter, are covered the principles which are basic to understanding properly the issues involved in the analysis of management data. This chapter cannot constitute the depth which goes into a specialized econometric book. It is, however, designed to provide the elements of econometric theory essential for a researcher to develop and evaluate regression models. Multiple regression is not a multivariate technique in a strict sense in that a single variable is the focus of the analysis: a single dependent variable. Nevertheless, the multivariate normal distribution is involved in the distribution of the error term which, combined with the fact that there are multiple independent or predictor variables, leads to considering simple multiple regression within the domain of multivariate data analysis techniques. The first section of this chapter presents the basic linear model with inferences obtained through the estimation of the model parameters. The second section discusses an important aspect of data analysis, especially in the context of testing contingency theories the issue of heterogeneity of coefficients. While many other econometric issues remain, such as autocorrelation or multicollinearity, the reader is referred to specialized books for these topics.

4.1

Statistical Inference: Least Squares and Maximum Likelihood

The linear model is first presented with its basic assumptions. Then, point estimates using the least squares criterion are derived, followed by the maximum likelihood estimation. Finally, the properties of these estimators are discussed.

4.1.1

The Linear Statistical Model

The dependent variable variables:

is modeled as a linear function of K independent

where T = number of observations (for example T periods), X = matrix of K independent variables, = vector of K weights applied to each independent variable k, y = vector of the dependent variable for t = 1 to T, e = vector of residuals corresponding to a unique aspect of y which is not explained by X.

56

Statistical Analysis of Management Data

It should be noted that X is given, fixed, observed data. X is, in fact, not only observable, but is also measured without error (the case of measurement error is discussed in Chapter 8). We assume that X is correctly specified. This means that X contains the proper variables explaining the dependent variable with the proper functional form (i.e., some of the variables expressed in X may have been transformed, for example by taking their logarithm). Finally, the first column of X is typically a vector where each element is 1. This means that the first element of the parameter vector is a parameter which corresponds to a constant term which applies equally to each value of the dependent variable from t = 1 to T. 4.1.1.1

Error structure

Some assumptions need to be made in order to be able to make some statistical inferences. Not all the assumptions below are used necessarily. In fact, in section 4.1.4.3, we identify which assumptions are necessary in order to be able to obtain the specific properties of the estimators. Because y and X are given data points and is the parameter vector on which we want to make inferences, the assumptions can only be on the unobserved factor e. Assumption 1: expected value of error term

Assumption 2: covariance matrix of error term Homoscedasticity Usually, each observation has an error term distributed with the same variance.

independently and identically

where I = identity matrix. This means that the variances for each observation t are the same and that they are uncorrelated. The unknown parameters which need to be estimated are: Heteroscedasticity More generally

Note that a covariance matrix, is a symmetric matrix. Heteroscedasticity occurs, therefore when This occurs if either the diagonal elements of the matrix are not identical (each error term has a different variance), and/or if its off-diagonal elements are different from zero.

Multiple Regression with a Single Dependent Variable

57

Assumption 3: normality of distribution The probability density function of the error vector can be written formally as per Equation (4.5) for the case of homoscedasticity or Equation (4.6) for the case of heteroscedasticity:

or

4.1.2

Point Estimation

Point estimates are inferences that can be made without the normality assumption of the distribution of the error term e. The problem can be defined as follows: to find a suitable function of the observed random variables y, given x, that will yield the “best” estimate of unknown parameters. We will restrict to the class that are linear functions of y.

The elements of the matrix A, are scalars that weight each observation; A is a summarizing operator. In order to solve the problem defined above, we need (1) to select a criterion, (2) to determine the A matrix and consequently and (3) to evaluate the sampling performance of the estimator. These three issues are discussed in the following sections. 4.1.2.1

OLS estimator

We now consider the case of homoscedasticity where

The criterion which is used to estimate the “best” parameter is to minimize the sum of squares residuals:

noting that is a scalar. This criterion is the least squares criterion and this problem is resolved by taking the derivative relative to the parameter vector setting it to zero and

58

Statistical Analysis of Management Data

solving that equation:

Note that the derivative in Equation (4.11) is obtained by using the following matrix derivative rules also found in the appendix: and

and especially:

Therefore, applying these rules to Equation (4.10), one obtains:

This assumes that can be inverted. If collinearity in the data exists, i.e., if a variable is a linear combination of a subset of the other x variables, the inverse does not exist (the determinant is zero). In a less strict case, multicollinearity can occur if the determinant of approches zero. The matrix may still be invertible and an estimate of will exist. We will briefly discuss the problem in subsection “computation of covariance matrix” of section 4.1.4.2. b is linear function of y:

where

4.1.2.2

GLS or Aitken estimator

In the general case of heteroscedasticity, the covariance matrix of the error term vector is positive definite symmetric:

The criterion is the quadratic form of the error terms weighted by the inverse of the covariance matrix. The rationale for that criterion is best understood

Multiple Regression with a Single Dependent Variable

59

in the case where is diagonal. In such a case, it can be easily seen that the observations with the largest variances are given a smaller weight than the others. The objective is then:

Minimizing the quadratic expression in Equation (4.20) is performed by solving the equation:

is still a linear function of y such as in Equation (4.14), but with the linear weights given by:

4.1.3

Maximum Likelihood Estimation

So far, the estimators which we have derived are point estimates. They do not allow the researcher to perform statistical tests of significance on the parameter vector In this section, we will derive the maximum likelihood estimators, which lead to distributional properties of the parameters. The problem is to find the value of the parameter which will maximize the probability of obtaining the observed sample. The assumption which is needed to derive the maximum likelihood estimator is the normal distribution of the error term: It is then possible to write the likelihood function which, for the homoscedastic case is:

or, for the case of heteroscedasticity:

60

Statistical Analysis of Management Data

We can then maximize the likelihood or, equivalently, its logarithm.

which is equivalent to minimizing the negative of that expression, i.e.,

This can be done by solving the derivative of Equation (4.28) relative to

which is simply the least square estimator. Similar computations lead to the maximum likelihood estimator in the case of heteroscedasticity which is identical to the generalized least squares estimator: We can now compute the maximum likelihood estimator of the variance by finding the value of that maximizes the likelihood or which minimizes the expression in Equation (4.28):

This is solved by setting the derivative relative to

to zero:

This results in:

which leads to the maximum likelihood estimator:

where is the vector of residuals obtained when using the maximum likelihood estimator of to predict y. The same computational approach can be done for the heteroscedastic case.

Multiple Regression with a Single Dependent Variable 4.1.4

61

Properties of Estimator

We have obtained estimators for the parameters and The next question is to find out how good they are. Two criteria are important for evaluating these parameters. Unbiasedness refers to the fact that on the average they are correct, i.e., on the average we obtain the true parameter. The second criterion concerns the fact that it should have the smallest possible variance. 4.1.4.1

Unbiasedness

Definition: b and and a fortiori the maximum likelihood estimators and are linear functions of random vector y. Consequently they are also random vectors with the following mean:

This proves the least square estimator is unbiased. Similarly for the generalized least squares estimator:

This means that on the average it is the true parameter; it is unbiased. 4.1.4.2

Best linear estimator

How do the linear rules above compare with other linear unbiased rules in terms of the precision, i.e., in terms of the covariance matrix. We want an estimator which has the smallest variance possible. This means that we need to compute the covariance matrix of the estimator and then, we will need to show that it has minimum variance. Computation of covariance matrix The covariance of the least squares estimator b is:

62

Statistical Analysis of Management Data

Therefore,

In the case of multicollinearity, is very large (because the determinants is close to zero). This means that the variance of the estimator will be very large. Consequently, multicollinearity results in parameter estimates which are unstable. The variance covariance matrix of the generalized least squares estimator is, following similar calculations:

BLUE (Best Linear Unbiased Estimator) Out of the class of linear unbiased rules, the OLS (or the GLS depending on the error term covariance structure) estimator is the best, i.e., provides minimum variance. We will do the proof with the OLS estimator when however, the proof is similar for the GLS estimator when The problem is equivalent to minimizing the variance of a linear combination of the k parameters for any linear combination. Let be a vector of constants.

The least squares estimator of is

The problem is therefore to find out if there exists another unbiased linear estimator which is better than the least squares estimator.

Multiple Regression with a Single Dependent Variable

63

An alternative linear estimator is

should be unbiased. This means that

By substitution of the expression of the estimator

For

to be unbiased, Equation (4.45) must be verified, i.e.,

This can only be true if and

What is the value of A which will minimize the variance of the estimator? The variance is: However,

Therefore,

The problem now is to minimize subject to the unbiasedness restrictions stated in Equations (4.49) and (4.50), i.e.,

64

Statistical Analysis of Management Data

This is a Lagrangian multiplier problem. The Lagrangian is

Therefore,

In addition,

Considering again the derivative relative to A given in Equation (4.55), i.e.,

replacing

by the expression obtained in Equation (4.56), we obtain

and, therefore,

Thus, the minimum variance linear unbiased estimator of replacing with the expression in Equation (4.59)

is obtained by

which is the one obtained from the ordinary least squares estimator:

We have just shown that the OLS estimator has minimum variance.

Multiple Regression with a Single Dependent Variable

65

4.1.4.3 Summary of properties

Not all three assumptions discussed in section 4.1.1 are needed for all the properties of the estimator. Unbiasedness only requires assumption #1. The computation of the variance and the BLUE property of the estimator only involve assumptions #1 and 2, and do not require the normal distributional assumption of the error term. Statistical tests about the significance of the parameters can only be performed with assumption #3 about the normal distribution of the error term. These properties are shown in Table 4.1. 4.2 Pooling Issues The pooling issues refer to the ability to pool together subsets of data. Therefore, this concerns the extent to which datasets are homogeneous or are generated by the same data generating function. This question can be addressed by testing whether the parameters of different subsets of data are the same or not. If the parameters are different, the objective may become, in a second stage, to develop models which contain variables explaining why these parameters differ. This would lead to varying parameter models which are outside the scope of this book. 4.2.1

Linear Restrictions

Let us write a linear model for two sets of data with respectively.

and

observations

where the y’s and the X’s represent the same variables in each subset of data. The subscripts in Equations (4.62) and (4.63) represent the two subsets of observations. For example, the dependent variable may be sales of a product

66

Statistical Analysis of Management Data

and X may contain a vector of 1’s for an intercept and the price of the product. The subscript can represent the country (countries 1 and 2 in this case). There would be time periods of observations in country 1 and periods in country 2. Assembling the two data sets together gives:

or where

can also be written as

or

which can also be written as: where R = [1 – 1]. This can be generalized to more than two subsets of data. This linear restriction can also be represented by the model

or Let RRSS be the restricted residual sum of squares coming from Equation (4.68) and URSS be the unrestricted residual sum of squares coming from Equation (4.64) or obtained by summing up the residual sum of squares of each equation estimated separately. Each one is distributed as a chi-square:

The test involves checking if the fit is significantly worse by imposing the constraint on the parameters. Therefore, a test of the restriction that the coefficients from the two data sets are equal is given by the following F test, which compares the residual sum of squares after corrections for differences in degrees of freedom:

This test necessitates that the number of observations in each set is greater than the number of parameters to have sufficient degrees of freedom. Otherwise, the unrestricted model cannot be estimated. If it is still possible

Multiple Regression with a Single Dependent Variable

67

to test that the observations are generated by the same model as the one used for the observations. The model is first estimated using only the observations from the first set of data, as in Equation (4.62). The residual sum of squares for these observations is Then, the pooled model is estimated as in Equation (4.68) to obtain the residual sum of squares RRSS. The two residual sums of squares and RRSS have independent chi squared distributions, each with respectively and degrees of freedom. The test of homogeneity of coefficients is therefore obtained from the significance of the difference between the two residual sums of squares:

Therefore, the test considers the F distribution:

4.2.2

Pooling Tests and Dummy Variable Models

In this section, we assume that there are multiple firms, individuals or territories. There are T observations for each of these N firms, individuals or territories. We can write the equation for a single observation The subscripts i and t indicates that the observations vary along two dimensions, for example individuals (i) and time (t). For example, represents sales in a district in a given month. can be expressed as a linear function of factors measured in this same territory at the same time period:

represents the intercept for observation i. This can be expressed in terms of an individual difference from a mean value of the intercept across all observations: which, when inserted into Equation (4.72) gives:

Depending on the nature of the variable model or an error component model.

the model is a dummy variable

68

Statistical Analysis of Management Data

If is fixed, then it is a dummy variable or covariance model. If is random, we would be facing an error component model. In this section, we consider the dummy variable model (i.e., is fixed). Model with constant slope coefficients and an intercept that varies over individuals. The dummy variable model can be represented for all the T observations in a given territory i as:

where

This is identical to creating a dummy variable for each observation if i = k and 0 otherwise. Equation (4.72) or (4.74) can be rewritten as:

where:

We can then form a vector of dummy variables for each territory Each of these dummy variables vector has T rows (T × 1) where each row is a 1. Then the full data can be expressed as:

Let us denote the residual sum of squares obtained from least squares estimation of Equation (4.77). This indicates that the model is Partially Restricted (PR) on the slopes which are assumed to be equal.

Multiple Regression with a Single Dependent Variable

69

The model with equal intercepts and different slopes is:

Let us denote the residual sum of squares obtained form the least square estimation of Equation (4.78). This indicates a Partial Restriction on the intercepts which are assumed to be the same. The model with complete restriction that the intercepts and slopes are equal is given by:

This equation is the completely restricted case where intercepts and slopes are assumed to be equal. This results in the residual sum of squares CRSS. Finally, the completely unrestricted model is one where slopes and intercepts are different. This model is estimated by running N separate regressions, for each individual or territory. The completely unrestricted residual sum of squares is CUSS. Homogeneity of intercepts and/or slopes can be tested using F tests based on the comparison of restricted and unrestricted residual sum of squares. The next section discusses the strategies for such pooling tests. Note that in all cases, the homogeneity along the second dimension is assumed. For example, homogeneity across time periods is assumed and pooling tests are performed across sections (i.e., firms, territories or individuals for example). 4.2.3

Strategy for Pooling Tests

The strategies follow from decomposing the tests over intercept and slopes. The process follows the one depicted in Figure 4.1. The first test consists of an overall test of homogeneity of intercept and slopes. For that purpose, the residual sum of squares from the completely unrestricted model (CUSS) is compared to the partially restricted model where intercept and slopes are restricted to be the same (CRSS). A failure to reject

70

Statistical Analysis of Management Data

Multiple Regression with a Single Dependent Variable

71

this test indicates that the intercept and slopes are all the same across sections. No more test is needed. In the case of rejection of the equality of intercepts and slopes, we now must test whether the difference comes from the intercept only, the slope only or both. Then another test is now performed to check for the equality of the slopes. For that purpose, we now compare the residual sum of squares from the completely unrestricted model (CUSS) with the residual sum of squares obtained from constraining the slopes to be equal A failure to reject the difference between these two models indicates that the slopes are equal. Because the slopes are equal but the full restriction leads to significant differences, one must conclude that the intercept is different across sections. If we reject the hypothesis of equal slopes, the slopes are different, in which case we must still find out if the intercept of the cross sections are the same or not. Therefore, a third test is performed where we now compare the completely unrestricted residual sum of squares (CUSS) with the residual sum of squares of the model with the restriction that the intercept is the same across sections A failure to reject the hypothesis indicates that slopes are the only source of heterogeneity (the intercept is the same across sections). A rejection of the test indicates that both intercept and slopes are different across sections. In this case, we started to check the source of heterogeneity by restricting the slopes and checking if the slopes were statistically different or not across sections. Instead, we could have first restricted the intercept, i.e., we could have tested for the homogeneity of the intercept first. If the hypothesis were rejected, we would then have tested for the homogeneity of slopes. This is the second line of tests shown in Figure 4.1. 4.3

Examples of Linear Model Estimation with SAS

Let us consider an example where the data set consists of the market share of four brands during seven periods. This market share is predicted by two variables, the percentage of distribution outlets carrying the brand during each period and the price charged for each brand during the period. Figure 4.2 shows an example of a SAS file to run a regression with such data. The data are first read: period (period), brand number (brandno), market share (ms), distribution (dist) and price (price). The variables are then transformed to obtain their logarithms so that the coefficients correspond to sensitivity parameters. Dummy variables for each brand except the first one are created. These will be used for estimating a model with different intercept for each brand. They are also used to compute new variables created for distribution and price for each brand. Three models are estimated as per the SAS file shown in Figure 4.2. The SAS procedure REG is first called. Then a model statement indicates the

72

Statistical Analysis of Management Data

model specification with the dependent variable on the left side of the equal sign and the list of independent variables on the right side. The first model is the completely unrestricted model where each brand has different intercept and slopes. A second model statement is used for the completely restricted model (same intercept and slopes for all the brands). Finally, the third model statement corresponds to the partially restricted model where each brand has a different intercept but the same distribution and price parameters. The output is shown in Figure 4.3. From the output, the residual sums of squares for the completely unrestricted model appears in the first model (i.e., CUSS = 0.14833). The degrees of freedom for this model is the number of observations (28 which follows from 4 brands with each 7 periods of data) minus the number of parameters (12), that is 16 degrees of freedom. The second model shows the completely restricted case where all intercepts are the same and the slopes are the same as well. There are 3 parameters estimated and the CRSS is 46.3733. The third model has a different intercept for each brand but the same slopes. Therefore, 6 parameters are estimated and the is 0.19812.

Multiple Regression with a Single Dependent Variable

73

74

Statistical Analysis of Management Data

Tests of poolability can then be performed following the discussion in section 4.2. The test for complete homogeneity is given by the statistic:

Checking on the table for the F distribution with 9 and 16 degrees of freedom, the difference is clearly significant and the hypothesis of complete homogeneity is clearly rejected. We then proceed with testing for the homogeneity of slopes. We therefore compare the completely unrestricted model with the model where the slopes are restricted to be equal, which corresponds to the specification of model 3. There are 6 parameters and the residual sum of squares is 0.19812. The test is, therefore,

Comparing this statistic with the critical value of F with 6 and 16 degrees of freedom, it is clear that the constraint does not imply a significantly worse fit. Consequently, we can conclude that the parameters of the distribution and price variables are homogeneous across the brands. However, each brand has a separate intercept.

Multiple Regression with a Single Dependent Variable

75

4.4 Assignment Two data sets are available which contain information about a market in which multiple brands compete in an industry composed of five market segments. The full description of the data is given in Appendix C. The PANEL.CSV data set contains information at the segment level while the INDUP.CSV data set provides information at the industry level. The file ASSIGN4.SAS in Figure 4.4 is a SAS file which reads both data sets (INDUP.CSV and PANEL.CSV) and merges the two files.

76

Statistical Analysis of Management Data

The assignment consists in developing a model using cross sections and time series data. For example, it is possible to model sales for each brand as a function of the price and the advertising for the brand, sales force sizes, etc. Regardless of the model, you need to test whether the intercepts and slopes are homogenous. As another example, lets say you decide to model the awareness of each brand as a function of the awareness in the prior period and of the brand advertising of the current period. You may want to test if the process of awareness development is the same across brands. References Basic Technical Readings Chow, G. C. (1960), “Tests of Equality Between Subsets of Coefficients in Two Linear Regression,” Econometrica, 591–605. Fuller, W. A. and G. E. Battese (1973), “Transformation for Estimation of Linear Models with Nested Error Structure,” Journal of the American Statistical Association, 68, 343 (September), 626–632. Maddala, G. S. (1971), “The Use of Variance Component Models in Pooling Cross Section and Time Series Data,” Econometrica, 39, 2 (March), 341–358. Mundlack, Y. (1978), “On the Pooling of Time Series and Cross Section Data,” Econometrica, 46, 69–85. Nerlove, M. (1971), “Further Evidence on the Estimation of Dynamic Economic Relations from a Time Series of Cross Sections,” Econometrica, 39, 2 (March), 359–382.

Multiple Regression with a Single Dependent Variable

77

Application Readings Bass, F. M. and R. P. Leone (1983), “Temporal Aggregation, the Data Interval Bias, and Empirical Estimation of Bimonthly Relations From Annual Data,” Management Science, 29, 1 (January), 1–11. Bass, F. M. and D. R. Wittink (1975), “Pooling Issues and Methods in Regression Analysis With Examples in Marketing Research,” Journal of Marketing Research, 12, 4 (November), 414–425. Bass, F. M., P. Cattin and D. R. Wittink (1978), “Firm Effects and Industry Effects in the Analysis of Market Structure and Profitability,” Journal of Marketing Research, 15, 3–. Bemmaor, A. C. (1984), “Testing Alternative Econometric Models on the Existence of Advertising Threshold Effect,” Journal of Marketing Research, 21,3 (August), 298–308. Bowman, D. and H. Gatignon (1996), “Order of Entry as a Moderator of the Effect of the Marketing Mix on Market Share,” Marketing Science, 15, 3, 222–242. Gatignon, H. (1984), “Competition as a Moderator of the Effect of Advertising on Sales,” Journal of Marketing Research, 21, 4 (November), 387–398. Gatignon, H. and P. V. Abeele (1997), “Explaining Cross-Country Differences in Price and Distribution Effectiveness,” Working Paper, INSEAD. Gatignon, H. and D. M. Hanssens (1987), “Modeling Marketing Interactions with Application to Salesforce Effectiveness,” Journal of Marketing Research, 24, 3 (August), 247–257. Gatignon, H., J. Eliashberg and T. S. Robertson (1989), “Modeling Multinational Diffusion Patterns: An Efficient Methodology,” Marketing Science, 8, 3 (Summer), 231–247. Gatignon, H., T. S. Robertson and A. J. Fein (1997), “Incumbent Defense Strategies against New Product Entry,” International Journal of Research in Marketing, 14, 163–176. Gatignon, H., B. A. Weitz and P. Bansal (1989), “Brand Introduction Strategies and Competitive Environments,” Journal of Marketing Research, 27, 4 (November), 390–401. Hatten, K. J. and D. Schendel (1977), “Heterogeneity within an Industry: Firm Conduct in the U.S. Brewing Industry, 1952–71,” Strategic Management Journal, 26, 2, 97–113. Jacobson, R. and D. A. Aaker (1985), “Is Market Share All that it’s Cracked Up to Be?,” Journal of Marketing, 49 (Fall), 11–22. Johar, G. V., K. Jedidi and J. Jacoby (1997), “A Varying-Parameter Averaging Model of On-line Brand Evaluations,” Journal of Consumer Research, 24, September, 232–247. Lambin, J.-J. (1970), “Optimal Allocation of Competitive Marketing Efforts: An Empirical Study,” Journal of Business, 43, 4 (October), 468–484. Miller, C. E., J. Reardon and D. E. McCorkle (1999), “The Effects of Competition on Retail Structure: An Examination of Intratype, Intertype, and Intercategory Competition,” Journal of Marketing, 63, 4 (October), 107–120. Montgomery, D. B. and A. J. Silk (1972), “Estimating Dynamic Effects of Market Communications Expenditures,” Management Science, 18, 10 (June), B485–501. Naert, P. and A. Bultez (1973), “Logically Consistent Market Share Models,” Journal of Marketing Research, 10 (August), 334–340. Parson, L. J. (1974), “An Econometric Analysis of Advertising, Retail Availability and Sales of a New Brand,” Management Science, 20, 6 (February), 938–947. Parson, L. J. (1975), “The Product Life Cycle and Time Varying Advertising Elasticities,” Journal of Marketing Research, 12, 3 (November), 476–480. Robinson, W. T. (1968), “Marketing Mix Reactions to Entry,” Marketing Science, 7, 4 (Fall), 368–385. Robinson, W. T. (1988), “Sources of Market Pioneer Advantages: The Case of Industrial Goods Industries,” Journal of Marketing Research, 25, 1 (February), 87–94. Robinson, W. T. and C. Fornell (1985), “Sources of Market Pioneer Advantages in Consumer Goods Industries,” Journal of Marketing Research, 22, 3 (August), 305–317. Steenkamp, J.-B. E.M., F. t. Hofstede, et al. (1999), “A Cross-National Investigation into the Individual and National Cultural Antecedents of Consumer Innovativeness,” Journal of Marketing 63, April, 55–69. Urban, G. L., T. Carter, S. Gaskin and Z. Mucha (1986), “Market Share Rewards to Pioneering Brands: An Empirical Analysis and Strategic Implications,” Management Science, 32 (June), 645–659.

This page intentionally left blank

5. System of Equations

In this chapter we consider the case where several dependent variables are explained by linear relationships with other variables. Independent analysis of each relationship by Ordinary Least Squares could result in incorrect statistical inferences either because the estimation is not efficient (a simultaneous consideration of all the explained variables may lead to more efficient estimators for the parameters) or may be biased in cases where the dependent variables influence each other. In the first section, a model of Seemingly Unrelated Regression is presented. In the second section, we discuss the estimation of simultaneous relationships between dependent or endogenous variables. Finally, in section three, we discuss the issue of identification when systems of equations are involved.

5.1

Seemingly Unrelated Regression (SUR)

The case of Seemingly Unrelated Regression occurs when several dependent variables are expressed as a linear function of explanatory variables, leading to multiple equations with error terms which may not be independent of each other. Therefore, each equation appears unrelated to the other. However, they are in fact linked by the error terms, which leads to a disturbance-related set of equations. We will first present the model. Then, we will derive the proper efficient estimator for the parameters and, finally, we will discuss the particular case when the predictor variables are the same in each equation. 5.1.1

Set of Equations with Contemporaneously Correlated Disturbances

Let us consider time series of M cross sections. Each cross section i presents T observations, usually over time, although t could represent individuals for which M characteristics are modeled. Therefore, for each cross section, the vector of dependent variables has T observations (the vector is dimensioned T × 1). In this equation for the i-th cross section, there are predictor variables. A priori, the variables explaining a dependent variable are different for each cross section or variable i. Consequently, the matrix contains T rows and columns. The linear equation for each cross section can, therefore, be represented by Equation (5.1):

80

Statistical Analysis of Management Data

Stacking all the cross sections together, the model for all cross sections can be expressed as:

where This can be written more compactly as:

The error terms have zero mean, variances which vary for each equation, i.e., and the covariance corresponding to the same time period t for each pair of cross section is All other covariances are zero. This can be expressed for each cross sectional vector of disturbances as

and

It may be useful to write the full expression for Equation (5.5) for two cross sections i and j:

Let be the contemporaneous covariance matrix, i.e., the matrix where each cell represents the covariance of the error term of two equations (cross sections) for the same t:

System of Equations

81

Consequently, using the Kronecker product, we can write the covariance matrix for the full set of cross sections and time series data:

The matrix expressed in Equation (5.8) can be visualized below:

5.1.2

Estimation

The structure of the covariance matrix of the error term is characteristic of heteroscedasticity. Consequently the Generalized Least Squares Estimator will be the Best Linear Unbiased Estimator:

However, from Equation (5.8) and using the property of the inverse of a Kronecker product of two matrices:

and, therefore,

This estimation only requires the inversion of an M × M matrix, the matrix of contemporaneous covariances.

82

Statistical Analysis of Management Data

The generalized Least Squares Estimator is unbiased:

Its variance-covariance matrix is:

In practice, the contemporaneous covariance matrix is, however, unknown. If it can be estimated by a consistent estimator, the Estimated Generalized Least Squares Estimator can be computed by replacing the contemporaneous covariance matrix in Equation (5.12) by its estimated value. is estimated by following the three steps below: Step 1: Ordinary Least Squares are performed on each equation separately to obtain the parameters for each equation or cross section i:

These OLS estimators are unbiased. Step 2: The residuals are computed:

Step 3: The contemporaneous covariance matrix can then be computed:

alternatively, the cross-product residuals can be divided by instead of T. The Estimated Generalized Least Squares Estimator is then found as:

It is then possible to compute the new residuals obtained from the EGLS estimation and recalculate an updated covariance matrix to find a new EGLS estimate. This iterative procedure converges to the maximum likelihood estimator. 5.1.3

Special Cases

There are two special cases where it can be demonstrated that the Generalized Least Squares Estimator obtained from the Seemingly Unrelated Regression is identical to the Ordinary Least Squares estimator obtained one

System of Equations

83

equation (cross section) over time. These two cases are when:

1. The independent variables in each equation are identical (i.e., same variables and same values):

2. The contemporaneous covariance matrix is diagonal, i.e., the errors across equations or cross sections are independent:

Consequently in both of these cases, there is no need to compute the covariance matrix.

5.2 A System of Simultaneous Equations 5.2.1

The Problem

Again, the problem consists in estimating several equations, each corresponding to a variable to be explained by explanatory variables. The difference with the prior situation for Seemingly Unrelated Regression is that the variables which are explained by the model can be an explanatory variable of another one, thereby creating an endogenous system. These variables are then called endogenous variables and the variables which are not explained by the system are exogenous variables. Therefore, we need to estimate the parameters of a system of N linear equations, where there are T observations for each equation. For one observation t:

The system of N equations for each t can therefore be expressed as:

where the matrices and B are matrices containing the parameters of all equations. For example in the case of two equations (i.e., two endogenous variables) and two exogenous variables:

84

Statistical Analysis of Management Data

This corresponds to the two equations:

In addition, the error terms have the following properties:

and the contemporaneous covariance matrix is the symmetric matrix:

while the non contemporaneous error terms are independent:

The reduced form can be obtained by post-multiplying Equation (5.21) by assuming the inverse exists:

or

where

The elements of the matrix are the parameters of the reduced form of the system of equations. The random term is distributed with the following mean and covariance:

Equation (5.28) represents a straightforward set of equations similar to those discussed in Section 1 for Seemingly Unrelated Regressions. We can always get estimates The issue is “can we go from to and i.e., is the knowledge about sufficient to enable us to make inferences about the individual coefficients of and Let us write the entire model represented by Equation (5.21) for the T observations (t = 1, . . . , T).

System of Equations

85

Let

and

Then, the system of equations is:

Similarly to what was done above by post-multiplying by the inverse of

or Because E[U] = 0, the Ordinary Least Squares Estimator of unbiased:

Therefore we can predict Why is this useful? Let us consider one equation (i = 1). Let and Then, the first equation can be represented by:

so that

are

86

Statistical Analysis of Management Data

Let

or Why can’t we estimate the parameter vector using Ordinary Least Squares? The reason is that the estimator would be biased due to the fact that and are correlated. This comes from the fact that and and are correlated due to Indeed, for example with two equations and one exogenous variable in each equation:

The covariance matrix between

and

Then, what can we do? We can predict

is:

from the reduced form which is:

This estimation is based on the Ordinary Least Squares estimates of the parameters which are obtained by regressing on the entire set of exogenous variables (not just the one in Equation 1, but in all the equations, as follows from Equation (5.42)). The OLS estimator is:

Therefore, the predicted values of

are given by:

Note that is not correlated with because the X’s are uncorrelated with and that is not correlated with because has been removed. Therefore, one can replace in Equation (5.38) by its predicted value

System of Equations 5.2.2

87

Two Stage Least Squares: 2SLS

This follows directly from the conclusion derived in the prior section. One can remove the bias introduced by the endogeneity of the dependent variables by first regressing separately each endogenous variable on the full set of exogenous variables and by using the estimated coefficients to predict each endogenous variable. In the second stage, each equation is estimated separately using the model as specified in each equation but replacing the actual values of the endogenous variables specified on the right hand side of the equation by its predicted values as computed from the first stage. More specifically: Stage 1: Regress using Ordinary Least Squares each y on all exogenous variables X

and compute the predicted endogenous variables Y:

Stage 2: Regress using Ordinary Least Squares each on the exogenous variables of that equation n and on the predicted endogenous as well as exogenous variables specified in that equation:

The parameters estimated

and

are unbiased.

However, because the non zero covariances diag the estimation does not provide efficient estimators. The purpose of the third stage in the Three Stage Least Square estimation method is to get efficient estimates, at least asymptotically.

5.2.3

Three Stage Least Squares: 3SLS

The first two stages are identical to those described above for the Two Stage Least Squares estimation. We now add the third stage: Stage 3: (i) Compute the residuals for each equation from the estimated coefficients obtained in the second stage:

88

Statistical Analysis of Management Data

(ii) Estimate the contemporaneous covariance matrix

where

(iii) Compute the Estimated Generalized Least Squares estimate similarly to the Seemingly Unrelated Regression case with the system of equations

5.3 5.3.1

Simultaneity and Identification

The Problem

The typical example used in economics to discuss the problem of identification concerns the supply and demand inter-relationships. While the curves of supply and demand in the Price–Quantity map can be represented as in Figure 5.1, we only observe and

System of Equations

89

The question consists, therefore, in determining how we can differentiate empirically between these two curves. A similar marketing example can be used to illustrate the problem with Sales and Advertising expenditures. While Sales are a function of Advertising expenditures, very often, Advertising budgets reflect the level of sales. This is especially an issue with cross sectional data. Therefore, we are facing the two functions:

The first equation is the market response function. The second equation is the marketing decision function. Fortunately, sales are not purely driven by advertising in most circumstances. Similarly, the decision regarding the advertising budget is a complex decision. The solution to the identification problem resides in specifying additional variables that will help differentiate the two curves. It is important to note that these additional variables (exogenous) in each equation must be different across equations; otherwise, the problem remains. 5.3.2 5.3.2.1

Order and Rank Conditions Order condition

If an equation n is identified, then the number of excluded variables in the equation n is at least equal to the number of equations minus 1 (i.e., N – 1). Therefore, checking for the order condition consists in making sure that each equation excludes on the right hand side at least N – 1 variables (exogenous or endogenous). This condition is necessary but not sufficient for the system of equations to be identified. 5.3.2.2

Rank condition

The rank condition provides necessary and sufficient conditions for identification. Recall the system of equations for a time period or cross section t:

We will use the example with two equations which, for a time period t can be written as:

90

Statistical Analysis of Management Data

or

It should be clear from Equation (5.56) that the two equations are indistinguishable. More generally, from Equation (5.54)

or

Let

Using again the case of two equations expressed in Equation (5.56):

Let be the row vector of zeros and ones which applied to the corresponding column vector defines a restriction imposed on equation n. For example, the restriction on equation 1 that can be expressed in a general way as It follows that by defining Indeed, we have then

By post-multiplying the restriction vector by the matrix A, the rank condition for the equation n to be identified is that the rank of this matrix is at least equal to the number of equations minus one. The equation is just identified if If the rank is less than N – 1, the equation is under-identified. If the rank is greater than N – 1, the equation is overidentified. The equation must be just or over-identified to be able to obtain

System of Equations

91

parameter estimates. For example:

if then is just identified.

Because N – 1 = 1 (N = 2), the first equation

5.4

Summary

In this chapter, we have presented the issue and estimation corresponding to multiple cases of simultaneity of variables. In fact, all the possible cases are embedded in the general case expressed in Equation (5.21). 5.4.1

Structure of

Matrix

If the matrix is diagonal, the system of equations is not simultaneous, except as expressed by the correlation of the error terms. In such a case, the model corresponds to the case of Seemingly Unrelated Regressions. If the matrix is not diagonal but triangular, this results in a system which is not truly simultaneous either. In such a case, a dependent variable may affect another one but not the other way around. The system is then recursive. The various estimations which are appropriate for each of these cases is summarized in Figure 5.2.

92

Statistical Analysis of Management Data

Following Figure 5.2, the estimation method depends on the model specification as reflected in the matrix discussed above and in the covariance structure of the error term 5.4.2

Structure of

Matrix

When is diagonal, the EGLS estimator provides an efficient estimator if the covariance matrix is not diagonal; otherwise, each equation can be estimated separately by OLS as the results are identical. If the covariance matrix is not diagonal, Seemingly Unrelated Regression must be used. If the matrix is triangular, i.e., the case of a recursive system, OLS estimation of each equation separately provides unbiased parameter estimates. However, in the case where the covariance matrix is not diagonal, the covariance structure must be taken into consideration and the EGLS obtained from the 3SLS procedure provides an efficient estimator. If is diagonal, there is no need to proceed with multiple stage estimation. Finally, if the system of equations is simultaneous, i.e., is neither diagonal or triangular, the OLS estimators would be biased. Therefore, depending on whether is diagonal or not, 2SLS or 3SLS should be used. This points out the importance of knowing the structure of the covariance matrix In most cases, it is an empirical question. Therefore, it is critical to estimate the covariance matrix, to report it and to use the estimator which is appropriate. This means that a test must be performed to check the structure of the error term covariance matrix 5.4.3

Test of Covariance Matrix

The test concerns the hypothesis that the correlation matrix of the error terms is the identity matrix (Morrison 1976):

where R is the correlation matrix computed from the covariance matrix Two statistical tests are possible. 5.4.3.1

Bartlett’s test

The following function of the determinant of the correlation matrix follows a chi square distribution with v degrees of freedom:

System of Equations

93

where T is the number of observations in each equation, N is the number of equations and i.e., the number of correlations in the correlation matrix. 5.4.3.2

Lawley’s approximation

The test statistic as expressed in Equation (5.64) can be approximated by

where only the upper half of the correlations are considered in the summation. 5.4.4

3SLS versus 2SLS

The EGLS estimator is only asymptotically more efficient than the OLS estimator. Consequently, in small samples, it is not clear what the property of the EGLS estimator is. Therefore, sometimes, when the sample size is small, it may be appropriate to report the 2SLS estimates instead of the 3SLS ones. 5.5 5.5.1

Examples Using SAS

Seemingly Unrelated Regression Example

In the example below, three characteristics of innovations developed by firms are modeled as a function of firm factors and industry characteristics. The SAS file will be presented without going into the details of the substantive content of the model in order to focus on the technical aspects. In Figure 5.3, it can be seen that after reading the file which contains the data, the variables are standardized and scales are built. The model is specified within the SAS procedure SYSLIN for systems of linear equations. The SUR statement following the PROC SYSLIN instruction indicates that the parameters will be estimated using Seemingly Unrelated Regression. The dependent variables concern the relative advantage of the innovation, the radicalness of the innovation and its relative cost. The model statements for each equation specifies the independent or predictor variables. Some variables are the same but others are different across equations. The same model can also be estimated with Iterative Seemingly Related Regression. The only difference with the single iteration SUR in the SAS instructions is that SUR is replaced with ITSUR (see Figure 5.4). The output of these two estimations is shown in Figures 5.5 and 5.6 respectively. First, in both cases the OLS estimation is performed for each equation separately and the results are printed in the output.

94

Statistical Analysis of Management Data

System of Equations

95

96

Statistical Analysis of Management Data

System of Equations

97

98

Statistical Analysis of Management Data

System of Equations

99

The correlations from the residuals estimated from the OLS estimates are then shown. A test should be performed to check that the correlation matrix is statistically significant from the identity matrix in order to detect whether it is useful to use the SUR estimator. Finally, the SUR estimates (i.e., the EGLS estimator) are provided for each equation. It can be seen from the output of the Iterative Seemingly Unrelated Regression that the steps are identical. The estimates reported are those obtained at the last step when convergence is achieved.

5.5.2

Two Stage Least Squares Example

In the example for two and three stage least squares, we now specify some endogeneity in the system in that some variables on the left hand side of an equation can also be found on the right hand side of another equation. In the example shown in Figure 5.7, the model definition shows that the variable “dadvl” is a predicted variable and is also found in the equation to predict “dcostl”. The endogenous variables are identified in a statement which lists the variable names after the identifier “ENDOGENOUS ”. The statement “INSTRUMENTS” lists all the exogenous variables in the system. These variables will be used in stage 1 of the estimation procedure to calculate the predicted values of the endogenous variables which will be used for the estimation in the second stage. The estimation method is simply indicated on the same procedure line by “2SLS”. The output shown in Figure 5.8 provides the estimates of the second stage for each equation.

100

Statistical Analysis of Management Data

System of Equations

5.5.3

101

Three Stage Least Squares Example

Similarly to the case of two stage least squares, the estimation method is simply indicated on the SYSLIN procedure line by “3SLS”, as shown in Figure 5.9. All other statements are identical to those for two stage least squares. The output for the 3SLS procedure provides first the estimates of the second stage for each equation (they are not shown in Figure 5.10 because they are identical to the SAS output shown in Figure 5.8. In Figure 5.10, however, the estimated correlation matrix of the error terms across equations are shown. A test of significance of the set of correlations can then be performed to know whether it can be useful to continue to the third stage. These third stage EGLS estimates are then provided in the SAS output.

102

Statistical Analysis of Management Data

System of Equations

5.6

103

Assignment

The data found in the files INDUP.CSV and PANEL.CSV which are described in the Appendix and for which Chapter 4 described how to read the data in SAS provide opportunities to apply the systems of equations discussed in this Chapter. The assignment consists simply in specifying a system of equations to be estimated via the proper estimation method, as presented in this chapter. The modeling exercise should include (1) a system of seemingly unrelated equations or a recursive system and (2) a model with simultaneous relationships.

104

Statistical Analysis of Management Data

Examples of such models can concern the following: 1. A model of the hierarchy of effects which consists in awareness, purchase intentions and sales. 2. A model of the sales or market share for multiple segments or for multiple brands. 3. A model of a market response function and marketing decision functions.

Proper justification of the estimation method used must be included (i.e., test of the covariance structure of the error terms). References Basic Technical Readings Dhrymes, P. J. (1978), Introductory Econometrics, New York, NY: Spriner-Verlag New York Inc. [Chapter 6]. Judge, G. G., W. E. Griffiths, R. C. Hill, H. Lutkepohl and T.-C. Lee (1985), The Theory and Practice of Econometrics New York, NY: John Wiley & Sons [Chapters 14 and 15]. Morrison, D. F. (1976), Multivariate Statistical Methods, New York, NY: McGraw-Hill Book Company. Parsons, L. J. and R. L. Schultz (1976), Marketing Models and Econometric Research, New York, NY: North Holland. Theil, H. (1971), Principles of Econometrics, John Wiley & Sons, Inc. [Chapters 9 and 10].

Application Readings Bass, F. M. (1969), “A Simultaneous Equation Regression Study of Advertising and Sales of Cigarettes,” Journal of Marketing Research, 6 (August), 291–300. Bayus, B. L. and W. P. Putsis, Jr. (1999), “Product Proliferation: An Empirical Analysis of Product Line Determinants and Market Outcomes,” Marketing Science, 18, 2, 137–153. Beckwith, N. E. (1972), “Multivariate Analysis Sales Responses of Competing Brands to Advertising,” Journal of Marketing Research, May, 168-. Cool, K. and I. Dierickx (1993), “Rivalry, Strategic Groups and Firm Profitability,” Strategic Management Journal, 14, 47–59. Cool, K. and D. Schendel (1988), “Performance Differences Among Strategic Group Members,” Strategic Management Journal, 9, 207–223. Gatignon, H. and J. -M. Xuereb (1997), “Strategic Orientation of the Firm and New Product Performance,” Journal of Marketing Research, 34, 1 (February), 77–90. Lambin, J.-J., P. Naert and A. Bultez (1975), “Optimal Marketing Behavior in Oligopoly,” European Economic Review, 6, 105–128. Metwally, M. M. (1978), “Escalation Tendencies of Advertising,” Oxford Bulletin of Statistics, 243–256. Norton, J. A. and F. M. Bass (1986), “Diffusion and Theory Model of Adoption and Substitution for Successive Generations of High-Technology Products,” Management Science, 33, 9 (September), 1069–1086. Parker, P. M. and L -H. Roller (1997), “Collusive Conduct in Duopolies: Multimarket Contact and CrossOwnership in the Mobile Telephone Industry,” RAND Journal of Economics 28,2 (Summer), 304–322. Reibstein, D. and H. Gatignon (1984), “Optimal Product Line Pricing: The Influence of Elasticities and Cross-Elasticities,” Journal ofMarketing Research, 21, 3 (August), 259–267. Schultz, R. L. (1971), “Market Measurement and Planning With a Simultaneous Equation Model,” Journal of Marketing Research, 8 (May), 153–164. Wildt, A. (1974), “Multifirm analysis of Competitive Decision Variables,” Journal of Marketing Research, 8 (May), 153–164.

6. Categorical Dependent Variables

In this chapter, we consider statistical models to analyze variables where the numbering does not have any meaning and, in particular, where there is no relationship between one level of the variable and another level. In these cases, we are typically trying to establish whether it is possible to explain with other variables the level observed of the criterion variable. The chapter is divided in two parts. The first part presents discriminant analysis, which is a traditional method in multivariate statistical analysis. The second part introduces quantal choice statistical models. The models are described, as well as their estimation. Their measures of fit are also discussed.

6.1

Discriminant Analysis

In presenting discriminant analysis, the discriminant criterion, which is at the basis of understanding of the methodology, is first introduced. Then the derivation and the explanation of the discriminant functions are provided. Finally, issues of classification and measures of fit are discussed.

6.1.1

The Discriminant Criterion

The objective in discriminant analysis is to determine a linear combination of a set of variables such that several group means will differ widely on this linear combination. Let p = number of independent variables, N = number of observations, number of observations for group j = 1 , . . . , K, K = number of groups, is the vector representing the values on p variables for one observation i,

is the vector of weights to be attributed to each of the p variables

to form a linear combination. Therefore, this linear combination is given by Equation (6.1):

We will assume that follows a multivariate normal distribution. It follows that each is normally distributed. The problem consists in finding v which is going to maximize the F -ratio for testing the significance of the overall difference among several group means on a single variable y.

106

Statistical Analysis of Management Data

This value F is given by the ratio of the between group variance to the pooled within group variance of the variable y:

where N = number of observations or individuals, K = number of groups, between group sum of squares, pooled within group sum of squares. In the case where there are only two groups and one single variable (K = 2, p = 1), it is the classic t test of a difference of two means. The problem, therefore, is to find the value of v which will maximize F. The ratio (K – 1)/(N – K) is a constant; therefore,

The pooled within group sum of squares is the sum over the groups ( j ) of the squares of the deviations of variable y from their group mean.

For each group j (where j = 1 , . . . , K), we can write the vector of the values obtained from the linear combination of the variables. This vector has elements corresponding to the number of observations in group j. Let

Then,

where

Therefore,

Categorical Dependent Variables

107

Let

Then,

Let

matrix composed of the vector of the means for group j of the p variables, repeated times and let

matrix composed of the vector of grand means (across all groups) repeated N times.

Therefore,

and consequently,

We can maximize (the discriminant criterion) by taking the first derivative relative to v and setting it equal to 0 (we use the matrix derivation rule A.2 in Appendix A: 2Av:

108

Statistical Analysis of Management Data

After replacing by Equation (6.12), it becomes:

which premultiplying by

which follows from Equation (6.11), in

gives:

Therefore, the solution for is given by the eigenvalues of solution for v is given by the corresponding eigenvectors of 6.1.2

and the

Discriminant Function

The matrix is not symmetric. In fact, there are K – 1 linearly independent rows in Consequently, the rank of B is K – 1. is of full rank (p); if it were singular, it could not be inverted. Therefore, the number of non-zero eigenvalues is the smaller of the rank of and of B, which is usually K – 1 (following from the fact that typically there are more variables than groups, i.e., K – 1 < p). This means that discriminant analysis provides K – 1 non-zero eigenvalues and K – 1 discriminant functions. The first discriminant function has the largest discriminant criterion value (eigenvalue), and each of the others has a conditionally maximal discriminant criterion value. The centroids for each group j consist of the mean value of y for the group for each of the K – 1 eigenvectors or discriminating functions:

where r represents the index for the rth eigenvalue and eigenvector:

These are the dimensions along which one can find the largest differences across groups.

Categorical Dependent Variables 6.1.2.1

109

Special case of K = 2

It is possible to estimate a multiple regression equation where the dependent variable is a dummy variable (0 for alternative 1 and 1 for the other alternative). Such a regression would yield weights for the independent variables which would be proportional to the discriminant weights. However, it is important to note that the t statistics should not be used. Indeed, the errors are not normally distributed with mean 0 and variance as will be demonstrated in the sections below. 6.1.3 6.1.3.1

Classification and Fit Classification

The issue we need to address now concerns how to classify the observations. A group prediction can be made, based on the value of the linear combination obtained from the first discriminant function: The group prediction then depends on the value obtained in Equation (6.20), relative to a critical value i.e., based on the sign of The rule can then be based on the distance from group means: assign observation i to the group to which it is closest (corrected for covariance). The mid points are then used as the critical values. For example in the two-group case, there is a single eigenvector:

The classification is based on the mid point:

Then the classification rule is: which is equivalent to defining

as:

Then, if then Group 1 else Group 2. Graphically, this is represented on Figure 6.1 below, where the dotted vertical line represents the

110

Statistical Analysis of Management Data

critical value appearing at the mid point between the mean of each of the two groups and As discussed above: or equivalently: For more than two groups (i.e., K > 2), similar concepts apply. Let

The rule consists of assigning i to group j if for all which means that is closer to k than to j. For example, for three groups: K = 3. We can compute (note that But because we do not need Then we can classify i as belonging to

For more than two groups, a plot of the centroids functions as axes can help to interpret them. 6.1.3.2

on the discriminant

Measures of fit

Fit measures are based on the ability of the discriminant functions to classify observations correctly. This information is contained in the classification table, as shown in Figure 6.2. Percent correctly classified The classification table is a K × K matrix which indicates the number or percentage of observations which are part of each group and which have been classified into that group (correctly classified) or into another group. It can be seen in Figure 6.2 that the diagonal cells represent the observations which are correctly classified. The percentage of correctly classified

Categorical Dependent Variables

111

observations can easily be computed as:

where number of observations actually in category i and predicted to be in category j, N = total number of observations. This measure of fit presents two problems: it uses the same N individuals for discrimination and prediction. This leads to an upward bias in the probability of classifying the observations correctly. A solution is to use a split sample for prediction. if the sample is not distributed evenly across the groups, i.e., the observed proportions are different across groups. Then by merely classifying all observations arbitrarily into the group with the highest proportion, one can get at least max classified correctly, where is the actual proportion of observations in Group j. Maximum chance criterion This last value, i.e., max is defined as the maximum chance criterion. Because it does not require any model to be able to arrive at such a rate of correct assignment to groups, this can be used as a minimum standard, and any model should be able to improve on this rate. Percent correctly classified by chance: the proportional chance criterion Assume 2 groups:

112

Statistical Analysis of Management Data

Let be the observed proportion of observations actually in group j, as defined earlier, and the proportion of observations classified in group j.

Let us assume that the discriminant function is meaningful. Then we want to classify in the same proportion as the actual groups. Under our decision rule, Therefore,

Equation (6.30) provides the formula for the proportional chance criterion. Tau statistic

The tau statistic involves the same rationale but standardizes the information:

where number of observations classified in group j, correctly classified observations.

number of

6.2 Quantal Choice Models In this section, we will introduce logit models of choice. Although probit models could also be discussed in this section, they will not be discussed because they follow the same rationale as for the logit model. We start by discussing the difficulties inherent in using the standard regression model with a categorical dependent variable, even a binomial one. Then we discuss methodologies which can be used to resolve some of those problems. We then present the logit model with two variants and explain the estimation of the logit model parameters. Finally, we present the various measures of fit. 6.2.1

The Difficulties of the Standard Regression Model with Categorical Dependent Variables

Let us assume the case of two groups. The variable representing the group assigment can take two values, 0 and 1:

Categorical Dependent Variables

113

This group assignment is made on the basis of a linear model:

Are the usual assumptions verified? 1. is

0? This would imply in this case that the error terms for each observation follow a specific random process. Indeed, from Equation (6.33) it follows that:

Consequently, the following distribution for the equality 0 be verified:

would be required so that

However, this is not generally the case, in part because

Therefore, the distribution is impossible. Hence, is biased. 2. is The second assumption is the homoscedasticity of the error terms. is distributed as a Bernoulli process.

This implies heteroscedasticity, and consequently Ordinary Least Squares are inefficient. 3. The range constraint problem: A third problem occurs due to the fact that the predicted values of the predicted variable can be outside the range of the theoretical values, which are either 0 or 1. 6.2.2

Transformational Logit

6.02.2.1

Resolving the Efficiency Problem

We may be able to solve the efficiency problem. Let us assume that the data can be grouped into K groups.

where the K groups correspond to “settings” of independent variables.

114

Statistical Analysis of Management Data

Let

where

is the number of 1’s in group j.

The model for a given group is:

For the entire K groups, the proportions are represented by

In Equation (6.40), the true proportion for group j is given by

Therefore, follows a binomial distribution:

The variance is obtained because

Therefore, dividing by

is such that:

Categorical Dependent Variables

115

Consequently, the covariance of the error term in Equation (6.41) is:

The generalized least squares estimator would be:

But is unknown. It can be replaced by a consistent estimator to obtain the Estimated Generalized Least Squares Estimator. Such an estimator of is:

where The ordinary Least Squares estimator b provides estimates for p which are consistent with the theoretical model specification. The Estimated Generalized Least Squares Estimator is:

Several problems remain: are (i) There is no guarantee that the predicted probabilities between 0 and 1: an empirical solution which has been recommended is to restrict the variance so that if set or is between 0 and 1. (ii) Even then, there is no guarantee that based on This points out the need to constrain the range of p to the interval [0,1]. 6.2.2.2

Resolving the Range Constraint Problem

We can solve the range constraint problem: Transformational logit. Let

It can be shown that:

116

Statistical Analysis of Management Data

Let

Then or, for the full sample:

Therefore, the Generalized Least Squares Estimator provides the minimum variance estimator: where

But is unknown. We can replace by in Equation (6.65) and obtain the Estimated Generalized Least Squares Estimator:

In practice, let us define

Therefore, we can perform a transformation of the right and of the left hand sides of the equation and obtain the Ordinary Least Squares of the transformed variables. Let

Categorical Dependent Variables

117

and consequently,

6.2.3

Conditional Logit Model

Let us consider an individual i considering a choice among K alternatives. Let us define the variable

Only one alternative can be chosen so that:

The likelihood function for all individuals is:

For the Multinomial logit model, if the unobserved utilities are a function of attributes and an error term that is distributed iid with the extreme value distribution (i.e., the cumulative distribution function is then the probability is defined as:

where represents the utility associated with alternative j for individual i. Two cases can be found depending on whether the explanatory variables determining the utility of the alternatives vary across alternatives or not. The first case of Conditional Logit – Case 1 (or discrete choice in LIMDEP) concerns the case where the variation in utilities of the alternatives comes from the differences in the explanatory variables but the marginal utilities are invariant. The second case of the Conditional Logit model – Case 2 is when the source of variation in the utilities of the alternatives come from the marginal utilities only.

118 6.2.3.1

Statistical Analysis of Management Data Conditional logit – case 1 (or discrete choice in LIMDEP)

The utility of an option varies because of different values of X’s (e.g., attribute values of a brand).

For identification, we set

or let us define

This demonstrates that no constant term can be estimated in this model; a constant term would be indeterminate because the intercept disappears in Equation (6.79). The model parameters are estimated by maximum likelihood. The likelihood for individual i is:

For the N observations, the likelihood is:

Categorical Dependent Variables

119

The optimization follows the iterative procedure described below: Let t = iteration number. The gradient at iteration t is

Let us further define:

The value of the parameters at the next iteration is given by Equation (6.87):

The parameter estimates are obtained by convergence when the gradient vector approaches zero. 6.2.3.2

Conditional logit – case 2

In this case, the utility of an option varies because of different values of the marginal utilities and the factors predicting the utilities are the same across options.

For identification, it is necessary to set The estimation of the model follows the procedure as in the prior case.

Taking the logarithms,

An iterative procedure similar to case 1 above is used to obtain the maximum likelihood estimates. The only difference compared with case 1 comes from the larger size of the vector of parameters. The vector of all coefficients at iteration t is the vector with (K – 1)p elements The interpretation is, therefore, somewhat more complex in the case 2 model. The marginal utilities due to the increase of a unit of an explanatory variable are different across alternatives. Therefore, for example, marginally, variable may contribute to the utility of alternative j but not significantly to the utility of alternative k.

120 6.2.4

Statistical Analysis of Management Data Fit Measures

The fit measures follow for the most part those used in discriminant analysis, which are based on the classification table. However, some additional measures are available because of the maximum likelihood estimation and its properties. 6.2.4.1

Classification table

These measures are the same as in discriminant analysis: Percentage of observations correctly classified Maximum chance criterion Proportional chance criterion Tau statistic 6.2.4.2

Statistics of fit

Because of the properties of the likelihood function, two statistics can be used to test the model. Log likelihood chi-square test The null model is that the marginal utilities, apart from the constant term, are zero: if n is the number of successes the binary case:

observed in T observations, e.g., in

where represents the maximum likelihood estimates of the parameters of the reduced model with no slopes and is the value of the likelihood function obtained with these parameter estimates. Taking the logarithm:

If is the value of the likelihood function estimated at the maximum likelihood estimate then,

Therefore, an obvious advantage of the logit model vis a vis discriminant analysis is that it offers the possibility of testing the significance of the model.

Categorical Dependent Variables

121

Likelihood ratio index or Based on the same properties, the following index can be used.

If the model is a perfect predictor in the sense that and when then,

when

When there is no improvement in fit due to the predictor variables, then,

6.3 6.3.1

Examples

Example of Discriminant Analysis Using SAS

In Figure 6.3, the SAS procedure “discrim” is used. The variables used to discriminate are listed after the “var” term and then the variable which contains the group numbering follows the term “class” to indicate that it is a categorical variable. The key sections of the SAS output are shown in Figure 6.4. The output of discriminant analysis clearly shows (Figure 6.4) the within group SSCP matrices (separately for each group), the pooled within SSCP matrix W, the Between group SSCP matrix B and the total sample SSCP matrix T. The raw (unstandardized) and standardized (correcting for the different units and variances of each of the variables) canonical coefficients,

122

Statistical Analysis of Management Data

Categorical Dependent Variables

123

124

Statistical Analysis of Management Data

Categorical Dependent Variables

125

126

Statistical Analysis of Management Data

Categorical Dependent Variables

127

that is the discriminant coefficients, are then listed. The raw coefficients indicate the weights to apply to the p variates in order to form the most discriminating linear function. In the example, In the particular case where only two groups are analyzed, a single discriminant function exists; there is only one eigen vector. The eigen vectors or discriminant functions discussed earlier are interpretable in a way such that a positive (negative) sign of the discriminant function coefficients (weights) indicates that the corresponding variable contributes positively (negatively) to the discriminant function. A comparison with the group means on the discriminant function indicates in what way the variates discriminate among the groups. For example, in Figure 6.4, Choice 1 has a higher (positive) mean value (0.142) on the discriminant function y (the mean for Choice 2 is negative, i.e., –0.488).

128

Statistical Analysis of Management Data

Therefore, the positive coefficient of DUNC means that the higher the demand uncertainty (the higher the value on DUNC), the higher the discriminant function and, consequently, the more likely choice 1 (internal development mode). On the opposite, because of the negative coefficient of TECHU, the higher the technological uncertainty, the more likely choice 2 of using an alliance. In addition, the absolute value of the standardized discriminant function coefficients (where the raw coefficients are multiplied by the standard deviation of the corresponding variables) reflect the contribution of the variables to that discriminant function so that a larger standardized weight indicates a bigger role of that variable in discriminating between the options. For example, the variable technology uncertainty (“techu”) appears the most discriminant variable (–0.75), followed closely by the variables “asc” (–0.71) and “grow” (0.69) although observations with higher values of growth (“grow”) are likely to belong to different groups from those with high ratings on “asc” and “techu” because of the opposite signs of these coefficients. Therefore, these standardized coefficients explain the contribution (extent and direction) of each variable for discriminating between the two groups. For two-group discriminant analysis, the interpretation of the discriminant function weights is relatively clear, as presented above. When there are more than two groups, each discriminant function represents different dimensions on which the discrimination between groups would occur. For example, the first discriminant function could discriminate between groups 1 and 3 versus group 2, and the second discriminant function could discriminate between groups 1 and 2 on the one hand and group 3 on the other hand. The interpretation in such cases requires the comparison of the group means on the discriminant function values (y). A plot of the group means or centroids on the discriminant functions as axes helps the interpretation of these discriminant functions which can be difficult. It is also very useful to analyze the profiles of each group in terms of the means of the predictor variables for each group. In Figure 6.4, a vector of coefficients for each group is printed under the heading of “linear discriminant function”. These are not, however, the discriminant functions discussed earlier; they are the classification functions. Indeed, in that particular example with two choices only, there could not be two discriminant functions. What the SAS output shows are the classification functions, which are the two components of equation (6.22) above, i.e., and The classification table is also shown in Figure 6.4. In this example, 62.58% of the observations in Group 1 were classified in the correct group and 73.3% for Group 2. 6.3.2

Example of Multinomial Logit – Case 1 Analysis Using LIMDEP

Figure 6.5 presents a typical input file using LIMDEP to estimate a logit model of the case 1 type. The data set used for this example, scanner.dat, has

Categorical Dependent Variables

129

the same structure as the data scan.dat described in Appendix C. The first part of the file defines the data variables and reads them from the data file. The specification of the analysis follows in the second part with the procedure “discrete choice”. The variables in the left hand side of the equation are then specified (purchase) following the code “lhs=”. Finally, the explanatory variables are listed after the code “rhs=” for the right hand side of the equation. It is important to note that in LIMDEP, the options must be coded from 0 to K – 1. The predicted variables in the example of Figure 6.5 consist of the price of each brand, any price cut applied to each transaction and whether the brand was on display on not. Each brand is also specified as having a different intrinsic preference or utility which is modeled as a different constant term with dummy variables (the reference where all brand dummies are zero correspond to private labels). Some heterogeneity in preferences across consumers is also captured by a loyalty measure representing past purchases of the brand. The LIMDEP output is shown in Figure 6.6. The output shown in Figure 6.6 should be self explanatory. The gradient is printed at each iteration until convergence is achieved. Then, the estimated parameters are listed with the usual statistics which enable the test of hypotheses and the computation of the fit statistics based on the likelihood function. The coefficients represent the marginal utility of each choice option (brand) of one additional unit of the corresponding variable. In the example in Figure 6.6, price has a significant negative impact while price cuts and being on display add to the brand utility.

130

6.3.3

Statistical Analysis of Management Data

Example of Multinomial Logit – Case 2 Analysis Using LOGIT.EXE

This program was written to estimate the parameters of a multinomial logit model. The model and the likelihood function are described in Gatignon, H. and E. Anderson (1988), “The Multinational Corporation’s Degree of Control Over Foreign Subsidiaries: An Empirical Test of a Transaction Cost Explanation,” Journal of Law, Economics and Organization, 4, 2 (Fall), 89–120. The program can be used for the binomial logit model specification as well. The multivariate logit program LOGIT.EXE runs on a personal computer under Windows. There are three files, in addition to the data file. These files,

Categorical Dependent Variables

131

including the data files, must be located in the same directory or folder. The three files are: LOGIT.EXE: the main source program in its compiled form which needs to be executed. LOGIT.PRM: the file containing the run parameters (see description below). LOGIT.RES: the file containing the output.

Executing the program Once the parameter file LOGIT.PRM has been created (using Notepad or a word processor and saved as text), the program can be executed by using Microsoft Explorer and clicking on the LOGIT.EXE file name. Again, this assumes that all the relevant files are in the same directory. The output is saved in the file LOGIT.RES and can be printed using Notepad or a word processor.

Description of parameter file Each line of the parameter file LOGIT.PRM is described in Figure 6.7 below on the right side of each line in italic characters. It should be noted that, contrary to LIMDEP, the options must be coded from 1 to K. The results are shown in the output file LOGIT.RES, as in Figure 6.8. The information provided is similar to the information provided in the LIMDEP output of the logit model – case 1. It should be pointed out that this example is a binomial logit case where the coefficients for option 1 are not shown since they are all zeros. Only the coefficients for option 2 and the corresponding statistics are shown. Finally, the output provides the classification table with the proportion of correctly classified observations (0.815 in Figure 6.8) and the value of the pseudo R-squared.

132

Statistical Analysis of Management Data

Categorical Dependent Variables 6.3.4

133

Example of Multinomial Logit – Case 2 Analysis Using LIMDEP

Figure 6.9 shows the LIMDEP file which estimates the same model as above. There are two aspects to pay particular attention to: 1. The choice variables should have a value of zero for the base case, up to the number of choice options minus one. In the example, the choice variable, which is the R&D mode is re-coded to take the value 0 or 1 dependent on whether the original variable read from the data file is 1 or 2. 2. The second point is that LIMDEP does not automatically estimate a constant term. Therefore, if one expects different proportions to be chosen for the same values of the independent variables, then the variable called “one” in LIMDEP serves to add the constant term.

It can be seen from the LIMDEP output, shown in Figure 6.10, that the results are the same as described previously, in terms of the parameter estimates and of the classification table. The information necessary to compute the likelihood ratio test are also given with the log-likelihood functions for the full model and for the restricted version (no slopes). The chi squared statistic is also provided. The pseudo R squared can be computed with this information as well.

134

Statistical Analysis of Management Data

Categorical Dependent Variables

135

6.4 Assignment Use SURVEY.ASC data to run a model where the dependent variable is a categorical scale (choose especially a variable with more than two categories). For example, you may want to address the following questions:

Can purchase process variables be explained by psychographics? Are demographics and/or psychographics determinants of media habits? Note that for these analyses, you can use Discriminant Analysis with SAS or the Multinomial logit – case 2 – model estimated using LOGIT.EXE or LIMDEP. In both cases (discriminant analysis and logit model), provide fit statistics in addition to the explanation of the coefficients. Compare the results of both analyses. Pay particular attention to the format for reading the variables in LIMDEP, as the windows version does not recognize format “i” for integers. Model the brand choice of orange juice using scanner data in the file SCAN.DAT (the description of the file can be found in Appendix C). Use LIMDEP to estimate the Multinomial logit – case 1 – models. You may want to consider the following ideas for possible analysis: What does the inclusion of the “loyalty” variable (i.e., a measure of crosssectional heterogeneity and nonstationarity) do to the brand choice model? What is the optimal value of the smoothing constant? Is it significantly different from 0.8? What do we gain, if anything, by separating price paid into its two components? Are there brand-specific price effects?

136

Statistical Analysis of Management Data

References Basic Technical Readings Maddala, G. S. (1983), Limited-dependent and Qualitative Variables in Econometrics, Cambridge: Cambridge University Press, [Chapters 3 and 4]. McFadden, D. (1974), “Conditional Logit Analysis of Qualitative Choice Behavior,” in P. Zarembka, ed., Frontiers in Econometrics, New York, NY: Academic Press. McFadden, D. (1980), “Econometric Models of Probabilistic Choice Among Products,” Journal of Business, 53, Sl3–. Morrison, D. G. (1969), “On the Interpretation of Discriminant Analysis,” Journal of Marketing Research, 6 (May), 156–163. Schmidt, P. and R. P. Strauss (1975), “The Prediction of Occupation Using Multiple Logit Models,” International Economic Review, 16, 2 (June), 471–486.

Application Readings Bruderl, J. and R. Schussler (1990), “Organizational Mortality: The Liabilities of Newness and Adolescence,” Administrative Science Quarterly, 35, 530–547. Corstjens, M. L. and D. A. Gautschi (1983), “Formal Choice Models in Marketing,” Marketing Science, 2, 1, 19. Fader, P. S. and J. M. Lattin (1993), “Accounting for Heterogeneity and Nonstationarity in a Cross-Sectional Model of Consumer Purchase Behavior,” Marketing Science, 12, 3, 304. Fader, P. S., J. M. Lattin and J. D. C. Little (1992), “Estimating Nonlinear Parameters in the Multinomial Logit Model,” Marketing Science, 11, 4, 372. Foekens, E. W., P. S. H. Leeflang and D. Wittink (1997), “Hierarchical versus Other Market Share Models for Markets with Many Items,” International Journal of Research in Marketing, 14, 359–378. Fotheringham, A. S. (1988), “Consumer Store Choice and Choice Set Definition,” Marketing Science, 7, 3 (Summer), 299–310. Gatignon, H. and E. Anderson (1988), “The Multinational Corporation’s Degree of Control Over Foreign Subsidiaries: An Empirical Test of a Transaction Cost Explanation,” Journal of Law, Economics and Organization, 4, 2 (Fall), 89–120. Gatignon, H. and D. J. Reibstein (1986), “Pooling Logit Models,” Journal of Marketing Research, 23, 3 (August), 281–285. Guadagni, P. M. and J. D. C. Little (1983), “A Logit Model Brand Choice Calibrated on Scanner Data,” Marketing Science, 2 (Summer), 203–238. Gupta, S. (1988), “Impact of Sales Promotions on When, What, and How Much to Buy,” Journal of Marketing Research, 25 (November), 342–355. Gupta, S., P. K. Chintagunta and D. R. Wittink (1997), “Household Heterogeneity and State Dependence in a Model of Purchase Strings: Empirical Results and Managerial Implications,” International Journal of Research in Marketing, 14, 341–357. Hardie, B. G. S., E. J. Johnson, and P. S. Fader (1992), “Modeling Loss Aversion and Reference Dependence Effects on Brand Choice,” Marketing Science, 12, 4, 378. Robertson, T. S. and H. Gatignon (1998), “Technology Development Mode: A Transaction Cost Conceptualization,” Strategic Management Journal, 19, 6, 515–532. Sinha, A. (2000), “Understanding Supermarket Competition Using Choice Maps,” Marketing Letters, 11,1,21–35. Tallman, S. B. (1991), “Strategic Management Models and Resource-Based Strategies Among MNEs in a Host Market,” Strategic Management Journal, 12, 69–82. Wiggins, R. R. and T. W. Ruefli (1995), “Necessary Conditions for the Predictive Validity of Strategic Groups: Analysis Without Reliance on Clustering Techniques,” Academy of Management Journal, 38, 6, 1635–1656. Yapa, L. S. and R. C. Mayfield (1978), “Non-Adoption of Innovations: Evidence from Discriminant Analysis,“ Economic Geography, 54, 2, 145–156.

7. Rank Ordered Data When the criterion variable is defined on an ordinal scale, the typical analyses based on correlations or covariances are not appropriate. The methods described in Chapter 6 do not use the ordered nature of the data and, consequently, do not use all the information available. In this chapter, we present methodologies that take the ordinal property of the dependent variable into account. A particular methodology which typically uses ordinal dependent variable is based on experimental designs to obtain preferences of respondents to different stimuli: conjoint analysis. We first discuss the methodology involved in conjoint analysis and the methods used to estimate the parameters of the conjoint models, i.e., monotone analysis of variance (MONANOVA). Then, we discuss a choice probability model which takes the ordinal property of the dependent variable into consideration, the ordered probit model. 7.1

Conjoint Analysis – MONANOVA

In the conjoint problem, preference responses to stimuli are obtained. These stimuli are designed to represent a combination of characteristics or attributes. Therefore, we start discussing the design itself which defines the independent or predictor variables and the manners in which the combination of attributes can be coded for analysis. 7.1.1

Effect Coding Versus Dummy Variable Coding

In a typical experimental setting, the independent variables which characterize the conditions of a cell or a stimulus are discrete categories or levels of attributes. For example the color of the packaging of a product is red or yellow. It can be ordered (for example a “low”, “medium” or “high” value) or not (e.g., colors). Each combination of level of all the attributes can correspond in principle to a stimulus, although responses to all the combinations may not be necessary. Two methods can be used to code these combinations of levels of attributes. Effect coding is the traditional method in experimental research using analyses of variance models. Dummy variables are typically used in regression analysis. We present each coding scheme and discuss the differences. The coding principle is best described by taking an example of a two by two factorial design. This means that there are two factors in the experiment, each with two levels. For example, the stimulus may or may not have property A and may or may not have property B. This is illustrated in Table 7.1.

138

Statistical Analysis of Management Data

This factorial design can easily be generalized to the design or any design m × n × … × k. In Table 7.1, the stimulus possesses the attribute A or not. If it does, the condition is noted as a, and if it does not, it is noted as The same two cases for attribute B are noted as b and The combinations of levels of the two attributes lead to the following cases: (1) = Treatment combination which consists of the 1st level of all factors, (a) = Treatment combination which consists of the 2nd level of the first factor and the 1st level of the second factor, (b) = Treatment combination which consists of the 1st level of the first factor and the 2nd level of the second factor, (ab) = Treatment combination which consists of the 2nd level of the two factors. These labels of each treatment condition are shown in each cell of the table describing the design in Table 7.1. Assuming that the various stimuli are evaluated on an intervally scaled response measure, the values also shown in each cell of Table 7.1 are the average ratings provided by respondents in each of these conditions. Assuming that the number of respondents in each cell are the same, one can derive the grand mean rating, the main effects of each attribute or factor and the specific incremental effect of the combination of A and B. The grand mean is the average value across the four cells:

The main effect of A is the average of the effect of the presence of A (i.e., the difference in the ratings whether A is present or not) across the two conditions determined by whether B is present or not. If B is present, the effect of A is (ab) – (b); if B is not present, it is (a) – (1), or:

Similarly, the main effect of B is the average of the effect of the presence of B (i.e., the difference in the ratings whether B is present or not) across the

Rank Ordered Data

139

two conditions determined by whether A is present or not. If A is present, the effect of B is (ab) – (a); if B is not present, it is (b) – (1), or:

The joint effect of A and B beyond the main effects of A and B is given by the difference between the value of the criterion variable when both effects are present and its value when none are present (i.e., (ab) – (1)), after removing the main effect of A (i.e., (a) – (1)) and the main effect of B (i.e., (b) – (1)):

Using the data in Table 7.1, (1) = 40.9, (ab) = 50.2, (a) = 47.8, (b) = 42.4. Therefore, using Equations (7.2) to (7.4):

These effects can simply be computed using a linear model where the independent variables are coded using a specific scheme. The coding scheme is different depending on whether the effects are coded directly (effect coding) or whether the levels are coded (dummy coding). 7.1.1.1

Effect coding

A variable will be created for each factor, for example for factor A and for factor B. We will first present the coding scheme with two levels and then when more than two levels are involved. Effect coding with two levels Let us assume a factor with two levels. The upper level will be coded “+1” and the lower level “–1”. Therefore, a stimulus (a cell) will be represented by the vector which for the four cells in Table 7.1 gives the following combinations:

A main effect model can be represented by the linear model:

140

Statistical Analysis of Management Data

The individual cells’ ratings can then be obtained by the combination of the values of and as indicated below:

For each cell, this leads to the equations:

The effects of each factor are therefore represented by the values of the

Effect coding with more than two levels When more than two levels are involved, the coding scheme depends on the assumptions made about the functional form of the relationship between the variable (factor) and the dependent variable. This issue obviously does not arise in the case of only two levels. We present below the case of three levels of a variable. The effects can be coded to reflect a linear relationship or a non-linear one. Linear effect coding Let us consider first the coding scheme for a linear effect. Such a coding is represented in Table 7.2. It can be seen that the difference between level one and level two is the same as the difference between level two and level three, that is one unit. The difference between level one and level three is twice the difference between level one an level two. Therefore, the effect is linear.

Rank Ordered Data

141

Non-linear effect coding The coding of non-linear effects varies depending on the functional form which the researcher wants to represent and test. Table 7.3 shows the coding scheme for a quadratic form. The shape of the function shows symmetry around level two and the values depend on the coefficient which multiplies this variable. Furthermore, a positive value of the coefficient would imply a decreasing and then increasing function and vice-versa for a negative coefficient. The coding scheme can become quite complex. For more than 3 levels, Table 7.4 provides the appropriate schemes. 7.1.1.2

Dummy variable

Dummy Coding corresponds to creating a variable (dummy variable) for each level of each factor minus one. Therefore, for a design where a factor has three levels, two variables are created: variable takes the value 0 for level one and level three, and 1 for level two and takes the value 0 for level one and level two, and 1 for level three. This implies that a separate coefficient will be estimated for each level, relative to the reference cell where all the dummy variables are 0. 7.1.1.3

Comparing effect coding and dummy coding

The two coding schemes do not give identical results because, from the presentation above, it is clear that effect coding places a restriction on the relationship which does not apply to dummy variable coding. Consequently, like any restricted form of a relationship compared to its unrestricted form, a test of the appropriateness of the restriction can be performed. The two approaches can consequently be combined to perform tests about the functional forms.

142

Statistical Analysis of Management Data

In summary, effect coding is appropriate when testing for the significance of the effect of a variable (conditionally on assuming a specific form of the relationship). Dummy coding is used to estimate and to test the effects of each level of a variable, independently of the other levels. 7.1.2

Design Programs

A particularity of conjoint analysis concerns the generation of the experimental design itself. Recently, several companies have developed PC based

Rank Ordered Data

143

software for generating stimuli reflecting the combination of the levels of attributes. Two such software packages are Conjoint Designer, by BrettonClark and Consurv, by IMS Inc. Each of these packages offer similar services where, once the attributes and their levels are determined, generate the combination of the attributes in the form of the description of the stimuli, enable the entry of the data by respondents and analyze the data.

7.1.3

Estimation of Part-worth Coefficients

In section 7.1.1, we have discussed one of the characteristics of conjoint analysis: the specific nature of the independent variables. The other characteristic of conjoint analysis concerns the rank ordered nature of the dependent variable. Although the term “conjoint” has recently been used in more broadly contexts, these two aspects were initially what distinguished conjoint analysis from other methodologies. MONANOVA was developed as an appropriate methodology for estimating the effects of variables using the rank ordered nature of the dependent variable. More recently, as conjoint studies developed successfully in industry, the simpler Ordinary Least Squares estimation has replaced the use of MONANOVA. This is due not only to the simplicity but also to two other factors: (1) the robustness of OLS which gives generally similar results to those obtained from MONANOVA and (2) the increased usage of ratings instead of rankings for the dependent variables. We first present MONANOVA and the estimation using PC-MDS. We then show how to perform OLS estimations using the SAS GLM procedure. 7.1.3.1

MONANOVA

Monotone Analysis of Variance is an estimation procedure based on an algorithm which transforms the dependent variable using a monotonic transformation so that the data can best be explained by a linear model of main effects of the independent variables or factors. More formally, let the data be represented by the set of values each corresponding to the evaluation of alternative j by individual i The data consists, therefore, for each individual of a table such as the one represented in Table 7.5.

144

Statistical Analysis of Management Data

The objective is, therefore, to estimate the main effects of each factor to fit best the relationship:

where f (·) is a monotonic transformation of the rank ordered dependent variable and and are the variables representing the main effects of the two factors using effects coding. The monotone transformations are performed using an algorithm to improve the fit. 7.1.3.2

OLS estimation

The GLM procedure found in SAS creates automatically the dummy variables which correspond to the design. By defining a variable as a discrete variable using the CLASS function, the levels of the variable are automatically generated with the proper dummy variables. The model is linear and the estimation follows the OLS estimation described in Chapter 4. It remains that MONANOVA is technically more appropriate when rank data is obtained and used as a dependent variable. This is particularly important for academic research where inappropriate methods should not be used, even if technically inappropriate methods provide generally robust results. Obviously, the use of ratings makes OLS a perfectly appropriate methodology. 7.2

Ordered Probit

Ordered probit modeling is a relatively recent approach to analysing rankordered dependent variables (McKelvey and Zavoina 1975). Let us assume that there exists an unobserved variable Y, which can be expressed as a linear function of a single predictor variable X. Furthermore, while the variable Y is not observed, only discrete levels of that variable can be observed (levels one, two and three). Figure 7.1 illustrates the case of a trichotomous dependent variable (observed variable) with a single independent variable. It is important to make the distinction between the theoretical dependent variable Y and the observed dependent variable Z, which, in the example of Figure 7.1 takes three possible values. The variable Y is an interval scale variable and, if we could observe it, it would fit a linear model The variable Z is ordinal, and generally presents M observed response categories The model of the unobserved dependent variable Y follows the usual linear model assumptions:

Rank Ordered Data

145

with We define M + 1 real numbers values:

with the following prespecified

These values are rank ordered such that Let us consider an individual observation i. The value of the dependent variable will be one if the underlying unobserved variable falls within the values of in the range of This can be expressed as:

Let us focus our attention on the interval in which the value of

falls.

We can replace the unobserved variable by the linear function of observed variables which determines it.

Subtracting the deterministic component from the boundaries, we obtain:

We can now standardize the values by dividing each element of the inequality by the standard deviation of the error term:

146

Statistical Analysis of Management Data

The central element is a random variable with the normal distribution:

We can therefore write the probability that this variable is within the range given by Equation (7.13) by subtracting the cumulative density functions at the upper and lower levels:

where

is the cumulative density function:

In order to identify the model, we need to impose the restrictions:

The first restriction has no consequence and the unit variance of the unobserved variable simply standardizes that variable. Consequently, Equation (7.15) reduces to:

The parameters which need to be estimated are:

This means that there are (K + M – 2) parameters to be estimated. The estimation is obtained by maximum likelihood. Let and, for simplification of the notation:

Then, the probability of

being in the interval

is

Rank Ordered Data

147

Consequently, the likelihood of observing all the values of Z for all the observations in the data set is:

The logarithm of the likelihood is:

The estimation problem consists in finding the values of the parameters which maximize the logarithm of the likelihood function subject to the inequality constraints about the values of i.e.,

One issue can be raised as, sometimes, it is not always clear whether the dependent variable is ordered or not. The question is then to know whether one is better off using ordered versus an unordered model. On the one hand, using an ordered model assumption when the true model is unordered creates a bias of the parameter estimates. On the other hand, using an unordered model when the true model is ordered does not create a bias but a loss of efficiency rather than consistency (Amemiya 1985, p. 293). Consequently, if the data is indeed ordered, the efficient and unbiased estimator will be provided by the Ordered model. Using an Unordered model may lead to parameters which are not significant but which would have been significant, had the most efficient model been used. Of course, this may not be an issue if all the parameters are significant. Using an ordered model if the data is not ordered is more dangerous because the parameter estimates are biased. Consequently, unless there is a strong theoretical reason for using an Ordered model, it is recommended to use a non-ordered model when the order property of the dependent variable is not clearly proven. 7.3 Examples 7.3.1 Example of MONANOVA Using PC-MDS

We will take the example of a 2 × 2 × 2 design where the data is as given in Table 7.6.

148

Statistical Analysis of Management Data

The MONANOVA program is ran by clicking on the monanova.exe file from Windows’ Exporer. The data as well as the information about the run are contained in an input file. An example is given in Figure 7.2. The first line shows the parameters of the problem, as shown in Table 7.7. The second line corresponds to the format in which the data can be read using FORTRAN conventions. The third line (and subsequent lines if there are more than one replication) corresponds to the data line(s). The data must be entered in a specific sequence. This sequence is best described through an example. In our 2 × 2 × 2 example, the indices of the x variable are such that the first index represents the level on the first factor, the second represents the level on the second factor and the third the level on the third factor. The sequence should then be as shown below: 111 112 121 122 211 212 221 222

The full input file is shown in Figure 7.2. The results of the MONANOVA analysis are shown in Figure 7.3.

Rank Ordered Data

149

150

Statistical Analysis of Management Data

Rank Ordered Data

151

The utilities for the levels within each factor are shown under the heading “UTILITIES OUTPUT FOR LEVELS WITHIN FACTORS.”

7.3.2

Example of Conjoint Analysis Using SAS

In the example below, data representing the ratings of different hypothetical schools are being used. The hypothetical schools were described in terms of (1) being either (a) not very or (b) very quantitative, (2) using methods of instructions characterized by (a) the case method, (b) half case and half lectures or (c) using only lectures, (3) the research reputation of the Faculty which can be (a) low, (b) moderate or (c) high, (4) the teaching reputation of the Faculty which can be also (a) low, (b) moderate or (c) high, and the overall prestige of the school as (a) one of the ivy league colleges, (b) a private school but not part of the ivy league and (c) a state school. The sample input file used with SAS is given in Figure 7.4. Figure 7.5 gives the output of such analysis. The tests of significance of each factor are performed and then the marginal means of the dependent variable is shown for each level of each factor, one at a time. The example also illustrates the test of some restrictions on the parameters such as for linear effects. 7.3.3

Example of Ordered Probit Analysis Using LIMDEP

The input file for LIMDEP which enables the estimation of an ordered probit model is straightforward (Figure 7.6). The only difference with the statements for a logit type model specification is the use of the command

152

Statistical Analysis of Management Data

Rank Ordered Data

153

154

Statistical Analysis of Management Data

“ORDERED”. It should be noted that the right hand side list of variables must include one. This particular example concerns the ranking of business schools as a function of ratings on the MBA program, the diversity of populations represented in the schools and the rating of research activities of the schools. Figure 7.7 shows the results of the analysis. Diversity appears insignificant but the rating of the MBA program as well as the rating of the school on R&D appear to strongly predict the overall ranking of the school.

7.4 Assignment 1. Decide on an issue to be analyzed with a conjoint study and gather data yourself on a few (10 to 20) individuals. Make sure that at least one of the factors has more than two levels. Investigate issues concerned with level of analysis and estimation procedures: Types of analysis: Aggregate analysis Individual level analysis Estimation: SAS GLM SAS with dummy variables SAS with effect coding MONANOVA 2. Using data from the SURVEY data, choose a rank ordered variable and develop a model to explain and predict this variable. Compare the multinomial logit model with the ordered logit or probit model. Use also a variable which is categorical and illustrate the problem of using an ordered logit or probit model when it is not appropriate.

Rank Ordered Data

155

156

Statistical Analysis of Management Data

References Basic Technical Readings Amemiya, T. (1985), Advanced Econometrics, Cambridge, MS: Harvard University Press, [Chapter 9]. Cattin, P., A. E. Gelfand and J. Danes (1983), “A Simple Bayesian Procedure for Estimation in a Conjoint Model,” Journal of Marketing Research, 20 (February), 29–35. Green, P. E. and V. R. Rao (1971), “Conjoint Measurement for Quantifying Judgmental Data,” Journal of Marketing Research, 8 (August), 355–363. Louviere, J. J. (1988), “Conjoint Analysis Modeling of Stated Preferences,” Journal of Transport Economics and Policy, 22, 1 (January), 93–119. McKelvey, R. D. and W. Zavoina (1975), “A Statistical Model for the Analysis of Ordinal Level Dependent Variables,” Journal of Mathematical Sociology, 4, 103–120.

Application Readings Beggs, S., S. Cardell and J. Hausman (1981), “Assessing The Potential Demand for Electric Cars,” Journal of Econometrics, 16, 1–19. Bowman, D. and H. Gatignon (1995), “Determinants of Competitor Response Time to a New Product Introduction,” Journal of Marketing Research, 32, 1 (February), 42–53. Bunch, D. S. and R. Smiley (1992), “Who Deters Entry? Evidence on the Use of Strategic Entry Deterrents,” The Review of Economics and Statistics, 74, 3 (August), 509–521. Chu, W. and E. Anderson (1992), “Capturing Ordinal Properties of Categorical Dependent Variables: A Review with Applications to Modes of Foreign Entry and Choice of Industrial Sales Force,” International Journal of Research in Marketing, 9, 149–160. Green, P. E. (1984), “Hybrid Models for Conjoint Analysis: An Expository Review,” Journal of Marketing Research, 21 (May), 155–169. Green, P. E. and V. Srinivasan (1978), “Conjoint Analysis in Consumer Research: Issues and Outlook,” Journal of Consumer Research, 5 (September), 103–123. Green, P. E. and V. Srinivasan (1990), “Conjoint Analysis in Marketing: New Developments With Applications for Research and Practice,” Journal of Marketing, October, 3–19. Green, P. E. and Y. Wind (1975), “A New Way to Measure Consumers’ Judgements,” Harvard Business Review, July–August, 107–117.

Rank Ordered Data

157

Green, P. E., A. M. Krieger and M. K. Agarwal (1991), “Adaptive Conjoint Analysis: Some Caveats and Suggestions,” Journal of Marketing Research, 28, May, 215–222. Jain, D. C., E. Muller, et al. (1999), “Pricing Patterns of Cellular Phones and Phonecalls: A Segment-Level Analysis,” Management Science 45, 2 (February), 131–141. Mahajan, V., P. E. Green and S. M. Goldberg (1982), “A Conjoint Model for Measuring Self- and CrossPrice/Demand Relationships,” Journal of Marketing Research, 19 (August), 334–342. Page, A. L. and H. F. Rosenbaum (1987), “Redesigning Product Lines with Conjoint Analysis: How Sunbeam Does it,” Journal of Product Innovation Management, 4, 120–137. Priem, R. L. (1992), “An Application of Metric Conjoint Analysis for the Evaluation of Top Managers’ Individual Strategic Decision Making Processes: A Research Note,” Strategic Management Journal, 13, 143–151. Rangaswami, A. and G. R. Shell (1997), “Using Computers to Realize Joint Gains in Negociations: Toward an “Electronic Bargaining Table”,” Management Science 43, 8 (August), 1147–1163. Srinivasan, V. and C. S. Park (1997), “Surprising Robustness of the Self-Explicated Approach to Customer Preference Structure Measurement”, Journal of Marketing Research, 34, 2 (May), 286–291. Wind, J., P. E. Green, D. Shifflet and M. Scarbrough (1989), “Courtyard by Marriott: Designing a Hotel Facility with Consumer-Based Marketing Models” Interfaces, 19 (January–February), 25–47.

This page intentionally left blank

8. Error in Variables – Analysis of Covariance Structure We will demonstrate that a bias is introduced when estimating the relationship between two variables measured with error. We will then present a methodology for estimating the parameters of structural relationships between variables which are not observed directly: the analysis of covariance structures. We will discuss especially the confirmatory factor analytic model, as presented in Chapter 3 and we will elaborate on the estimation of the measurement model parameters.

8.1 The Impact of Imperfect Measures In this section, we discuss the bias introduced by estimating a regression model with variables which are measured with error. 8.1.1

Effect of Errors-in-variables

Let us assume two variables, a dependent variable and an independent variable, and respectively, which are observed. However, these variables are imperfect measures of the true unobserved variables and The measurement models for both variables are expressed by the equations:

There exist a relationship, a structural relationship, between these two unobserved variables, as indicated by the equation below:

This equation can be expressed in terms of the observed variables by replacing the unobserved variables by their expression as a function of the observed variables obtained from Equations (8.1) and (8.2):

or, placing the random error terms at the end:

It should be noted that the error on the dependent variable y is similar to the error on the structural relationship. Indeed, if we had added an error term

160

Statistical Analysis of Management Data

to Equation (8.3), it would have been confounded with the measurement error of the dependent variable Because the variables are not observed, only the relationship between the observed variables can be estimated. This can be done by using the Ordinary Least Square estimator of the regression of on

The bias can be evaluated by taking the expectation of the OLS estimator:

where

Therefore, the bias is:

if x has mean of 0, the bias is

since

the bias can be expressed as

where is the signal to noise ratio. From Equation (8.10), we can not only assert that there is a bias but we can also indicate properties about this bias. Because the variances in the signal to noise ratio are positive this means that the bias is always negative (Equation (8.10) is always negative), i.e., the OLS estimates are under-estimated when using a predictor variable with error. This may lead to failing to reject the null hypothesis that the effect of the independent variable on the dependent variable is insignificant. As the signal to noise ratio increases, the bias decreases becomes smaller]. Therefore, we can summarize the results as follows: 1. We have found a lower bound for b. Indeed, we have shown that the OLS estimator is smaller than the true

Error in Variables – Analysis of Covariance Structure

161

2. Error in measurement of x attenuates the effect of x. 3. Error in measurement of y does not bias the effect of x (the measurement error is then confounded with the noise in the relationship between the independent and dependent variables). 8.1.2

Reversed Regression

Let us write the equation which expresses the independent variable function of the dependent variable

Or, for all the observations:

The OLS estimator of the parameter

is:

Let

If the variables are centered to zero mean:

However, from Equations (8.2) and (8.3):

Consequently, and

Therefore, Equation (8.15) can be expressed as

where

which is always positive.

as a

162

Statistical Analysis of Management Data

Because is positive, it follows that overestimates If we recall that the coefficient obtained from a direct regression (Equation (8.6)), which we may call

always under-estimates the true value of

we then have shown

that and provide bounds in the range where the true falls. Consequently, the choice of the dependent variable in a simple regression has nothing to do with causality. It follows from the analysis presented above that if is small, one should use reversed regression in Equation (8.19) is then close to 0 and the bias is small). If, however, is small, direct regression should be used because the bias in Equation (8.6) is then small. From this discussion, it follows that the researcher should select for the dependent variable, the variable with the largest measurement error. 8.1.3

Case with Multiple Independent Variables

The case where there are several independent variables is more complex. Let us consider Equation (8.20) where some variables are estimated without error, and others are estimated with measurement error:

In such cases, the direction of the bias is not easy to analyze. Some conclusions are possible, however, in the special case when only one of the independent variables is measured with error, i.e., is a single variable. Then, it can be shown that the bias can be expressed as follows:

where is the R-squared of the regression of the variable measured with error on those measured without error Because the ratio which multiplies in Equation (8.21) is always positive, the coefficient is, therefore, always under-estimated. It should be noted that having one of the independent variable measured with error does not affect only the estimation of the impact of that variable. It also affects the coefficients of the variables measured without error. Furthermore, both the overall F statistics and the individual coefficient variances are affected. The F Statistic is always understated. Therefore, we would expect to reject the models more often than we should. The impact on individual statistics is not as clear, however, as there is no unambiguous bias. This case of a single variable measured with error is, however, unusal. Most of the research in the social sciences involves the formation of scales that cannot be considered to be without measurement error. In such cases, the analysis shown in the first section of this chapter does not provide any guidance. The second section presents a methodology, the analysis of covariance structure, which takes care of the problems associated with measurement errors.

Error in Variables – Analysis of Covariance Structure

163

8.2 Analysis of Covariance Structures In the analysis of covariance structure, both the measurement errors and the structural relationships between the variables of interest are modeled.

8.2.1

Description of Model

We start with a system of simultaneous equations identical to the ones analyzed in Chapter 5:

where m = number of endogenous constructs, n = number of exogenous constructs, column vector of m endogenous constructs, column vector of n exogenous constructs, column vector of m disturbance terms, B = matrix of structural parameters of endogenous variables, matrix of structural parameters of exogenous variables. The endogenous constructs are represented by the vector and the exogenous ones by Equation (8.22) represents the structural relationships that exist among the constructs and with a random disturbance The diagonal elements of the matrix B are specified as being equal to one without affecting the generality of the model. The endogenous and exogenous constructs and are not observed but are, instead, measured with error using multiple items. Before defining the measurement models, we should note that these unobserved constructs are defined as centered with zero mean without any loss of generality. Like for the regression model, the error term is assumed to have zero mean:

In addition, the matrix of parameters B should be nonsingular. Let us now define the factor analytic measurement models. These are represented by Equations (8.25) and (8.26). There are p items or observable variables reflecting the m endogenous constructs and there are q items or observable variables reflecting the n exogenous constructs

where p = number of items measuring the m endogenous constructs, y = column vector of the p items or observable variables reflecting the endogenous constructs, matrix of factor loadings, column vector of measurement errors.

164

Statistical Analysis of Management Data

The elements of the matrix represent the factor loadings. Similarly for the measurement model of the exogenous constructs:

where q = number of items measuring the n exogenous constructs, x = column vector of the q items or observable variables reflecting the exogenous constructs, matrix of factor loadings, column vector of measurement errors. Furthermore, we can express the covariances of the latent variables and of the error terms according to Equations (8.27) through (8.30).

We can now write the expression of what would theoretically be the covariance matrix of all the observed variables (x and y), assuming the model expressed in the equations above. Let

The theoretical covariance matrix of z is

We will derive the expression of each of the four submatrices in Equation (8.32) with the following three blocks (the off-diagonal blocks are symmetric):

Error in Variables – Analysis of Covariance Structure

165

Equations (8.33) through (8.35) provide the information to complete the covariance matrix in Equation (8.32). In fact, the observed covariance matrix can be computed from the sample of observations:

8.2.2

Estimation

The estimation consists in finding the parameters of the model which will replicate as closely as possible the observed covariance matrix in Equation (8.36). For the maximum likelihood estimation, the comparison of the matrices S and is made through the following expression:

The expression F is minimized (note that the last term is a constant) by searching over values for each of the parameters. If the observed variables

are distributed as a multivariate normal distribution,

the parameter estimates that minimize the Equation (8.37) are the maximum likelihood estimates. There are distinct elements which constitute the data (this comes from half of the symmetric matrix to which one needs to add back half of the diagonal in order to count the variances of the variables themselves (i.e., [ (p + q ) × (p+q)/2+(p + q)/2]). Consequently, the number of degrees of freedom corresponds to the number of distinct data points as defined above minus the number of parameters in the model to estimate. An example will illustrate the model and the degrees of freedom. MacKenzie, Lutz and Belch (1986) compare several models of the role of attitude toward the ad on brand attitude and purchase intentions. Focusing on their dual mediation hypothesis model (DMH) which they found to be supported by the data, three types of cognitive responses to advertising (about

166

Statistical Analysis of Management Data

the ad execution, about the source and about repetition) are the three exogenous constructs explaining the attitude toward the ad. Attitude toward the ad, according to that DMH theory, affects the attitude towards the brand not only directly but also indirectly by affecting brand cognitions which, in turn, affect the attitude toward the brand. Purchase intentions are affected by the attitude towards the brand as well as directly by the attitude towards the ad. These relationships between the three exogenous constructs and these four endogenous constructs are drawn in Figure 8.1. These relationships can be expressed by the system of four equations:

or

In addition, Figure 8.1 indicates that the exogenous constructs are each measured by a single item, for for and for The attitude towards the ad is measured by two items and The attitude towards the brand and purchase intentions are both measured by three items, and for and and for Finally, the brand cognitions are measured by a single indicator The measurement model for the endogenous constructs can then be represented by Equation (8.40) and the measurement model for the exogenous constructs can be expressed by Equation (8.41):

and

Error in Variables – Analysis of Covariance Structure

167

It should be noted that some restrictions on the measurement model parameters must be made for identification purposes. For each construct, the unit or scale of measurement must be defined. This is accomplished by setting one of the lambdas for a given construct to one; the corresponding variable will then serve as the unit of reference for that construct. For example, we

Statistical Analysis of Management Data

168

can define Alternatively, especially in the case of confirmatory factor analysis, the variance of the constructs could be set to unity. We also need to impose some restrictions on some parameters in the cases where the constructs are measured by a single item. In such cases indeed, the loading parameter is set to one, as discussed above and the error term is necessarily equal to zero. This means that the variance of the error term of that measurement equation must be constrained to be zero. This is the case for the example with and Normally, the covariance matrices and are assumed to be diagonal. Exceptionally, a few of the correlations between error terms of measurement equations can be estimated. This was the case in the example reported above from MacKenzie, Lutz and Belch (1986). However, it should only be done with great care, as the interpretation may be difficult. The covariance matrix of the exogenous constructs can be symmetric or, with orthogonal factors it can be defined as diagonal with zero covariances. With orthogonal factors in the example, three variances must be estimated. Finally, the covariance matrix must be specified. It can be symmetric in the general case where the error terms of the structural equations are correlated. In this example, there would be four variances and six covariances to estimate. The matrix is often assumed to be diagonal, in which case only four parameters (four variances) need to be estimated. The equations described and the restrictions applied above indicate that 29 parameters must be estimated: there are 5 lambdas, 6 betas, 3 gammas, 8 thetas, 4 phis and 3 psis. Given that with twelve observed variables the covariance matrix consists of 78 different data points (i.e., (12 × 13)/2), this leaves 49 degrees of freedom.

8.2.3

Model Fit

The measure of the fit of the model to the data corresponds to the criterion which was minimized, i.e., a measure of the extent to which the model, given the best possible values of the parameters, can lead to a covariance matrix of the observed variables that is sufficiently similar to the actually observed covariance matrix. We first present and discuss the basic chi-squared test of the fit of the model. We then introduce a number of measures of fit that are typically reported and which alleviate the problems inherent to the chi-squared test. We finally discuss how modification indices can be used as diagnostics for model improvement. 8.2.3.1

Chi-square tests

Based on large sample distribution theory, (where N is the sample size used to generate the covariance matrix of the observed variables

Error in Variables – Analysis of Covariance Structure

169

and is the minimum value of the expression F as defined by Equation (8.37) is distributed as a chi-squared with the number of degrees of freedom corresponding to the number of data points minus the number of estimated parameters, as computed in the example above. If the value of is significantly greater than zero, the model is rejected; this means that the theoretical model is unable to generate data with a covariance matrix close enough to the one obtained from the actual data. One problem with that expression is due to the fact that it contains N, the sample size, in the numerator. This means that as the sample size increases, it becomes less likely that one will fail to reject the model. This is why several other measures of fit have been developped. They are discussed below. Another problem inherent to the methodology is the fact that the hypothesis which the researcher would like to get support for is the null hypothesis that there is no difference between the observed covariance matrix and the matrix that can be generated by the model. Failure to reject the hypothesis, and therefore “accepting” the model, can, therefore, be due to the lack of power of the test. This is clearly apparent in the issue mentionned above concerning the sample size. A small enough sample size can contribute to finding fitting models based on chi-squared tests. It should be noted that the sample size problem disappears when two nested models are compared. Indeed, the test of a restriction of a subset of the parameters implies the comparison of two of the measures of fit each distributed as a chi-squared. Consequently, the difference between the value of a restricted model and the unrestricted model, follows a chi-squared distribution with a number of degrees of freedom corresponding to the number of restrictions. This, in fact, is the basis for most of the goodness of fit measures that have been proposed.

8.2.3.2

Other goodness of fit measures

Of all the models possible, one extreme case is when there is no relationship postulated between any pair of the constructs. This model of statistical independence can be considered the null model. At the other extreme, it is possible to consider the model where all the relationships are estimated (a “saturated model”), in which case there are zero degrees of freedom. These two models provide the extreme values of the chi-squares and the difference between these chi-squares is the largest one possible: any other model which is a restricted version of the “saturated” model will have a better fit than the null model but worse than the saturated one. It is then possible to use an index of fit based on these ideas. One measure, the Bentler and Bonnet (1980) goodness-of-fit index (GFI) is the percentage improvement over the null model relative to the largest possible improvement that would be realized with a saturated model. This index is a pseudo R squared and is similar to an R squared in the sense that it varies between 0 and 1 and that approaching 1 means a better fit versus

170

Statistical Analysis of Management Data

approching 0 means a bad fit; however, it cannot be interpreted as a percentage of variance explained. Another GFI is provided in the LISREL output; it is a more direct measure of the fit between the theoretical and observed covariance matrices following from the fit criterion of Equation (8.37) and it is defined as

From this equation, it is clear that if the estimated and the observed variances are identical, the numerator of the expression substracted from 1 is 0 and, therefore, GFI = 1. To correct for the fact that the GFI is affected by the number of indicators, an adjusted Goodness of Fit Index (AGFI) is also proposed. This measure of fit corrects the GFI for the degrees of freedom, just like an adjusted R squared would in a regression context:

where t is the number of estimated parameters. As the number of estimated parameters increases, holding everything else constant, the adjusted GFI decreases. A threshold value of 0.9 (either for the GFI or AGFI) has become a norm for the acceptability of the model fit (Bagozzi and Yi 1988, Baumgartner and Homburg 1996, Kuester, Homburg and Robertson 1999). 8.2.3.3

Modification indices

The solution obtained for the parameter estimates uses the derivatives of the objective function relative to each parameter. This means that for a given solution, it is possible to know the direction in which a parameter should change in order to improve the fit and how steeply it should change. As a result, the modification indices indicate the expected gains in fit which would be obtained if a particular coefficient should become unconstrained (holding all other parameters fixed at their estimated value). Although not a substitute for theory, this modification index can be useful to analyze structural relationships and in particular to refine the correlational assumptions of random terms and for the modeling of control factors.

8.2.4

Test of Significance of Model Parameters

Because of the maximum likelihood properties of the estimates, the significance of each parameter can be tested using the standard t statistics formed by the ratio of the parameter estimate and its standard deviation.

Error in Variables – Analysis of Covariance Structure

8.2.5

171

Simultaneous Estimation of Measurement Model Parameters with Structural Relationship Parameters Versus Sequential Estimation

It can be noted that in the estimation method described above, the measurement model parameters are estimated at the same time as the structural model parameters. This means that the fit of the structural model and the structural model parameters are affected by the measurement model parameters. The motivation of the approach was to correct the bias produced by errors in measurement. However, the simultaneity of the estimation of all the parameters (measurement model and structural model) implies that a trade off is made between the values estimated for the measurement model and those for the structural model. In order to avoid this problem, it is a better practice to estimate first the measurement model and then estimate the structural model parameters in a fully specified model (i.e., with the measurement model) but where the parameters of the measurement model are fixed to the values estimated when this measurement model is estimated alone (Anderson and Gerbing 1988). This procedure does take into account the fact that the variables in the structural model are measured with error in order to estimate the structural model parameters, but does not let the estimation of the measurement model interfere with the estimation of the structural model and vice-versa.

8.2.6

Identification

As discussed earlier in Chapter 5, a model is identified if its parameters are identified, which means that there is only one set of values of the parameters that generate the covariance matrix. Although there is no general necessary and sufficient conditions for the general model discussed here to be identified, if the information matrix is not positive definite, the model is not identified. Furthermore, it appears logical that the structural model be identified on its own. The order and rank conditions presented in Chapter 5 should consequently be used to verify the identification of the structural relationships in an analysis of covariance structure model.

8.3 Examples

We now present examples of analysis of covariance structure using LISREL8 for Windows or AMOS. These examples include the test of a single factor analytic structure, the estimation of a factor analytic structure with two correlated factors and a full structural model with error in measurement.

172

8.3.1

Statistical Analysis of Management Data Example of Confirmatory Factor Analysis

The following example in Figure 8.2 shows the input file for LISREL8 for Windows: An exclamation mark indicates that what follows is a comment and is not part of the LISREL8 commands. Therefore, the first real input line in Figure 8.2 starts with DA which stands for data. On that line, NI indicates the number of input (observed) variables (6 in this example), MA = KM indicates the type of matrix to be modeled, KM for correlation or CV for covariance. The second line of input is used to specify how to read the data. RA indicates that the raw data will be read (from which the correlation matrix will be automatically computed) and FI = filename indicates the name of the file containing that data, where filename is the Windows file name including the full path. The third line, with LA, indicates that next come the labels of the indicator (input) variables. These are shown as Q5, Q7, etc. on the following line. The next line specifies the model, as indicated by the code MO at the beginning of that line. NX indicates the number of indicators corresponding to the exogenous constructs (here, there are six). NK stands for the number of ksi constructs (we have a unique factor in this example). PH = ST indicates that the covariance matrix phi is specified here as a standardized matrix, i.e., a correlation matrix with 1 ’s in the diagonal and 0’s off-diagonal. The covariance matrix of the measurement model error terms, theta delta, is specified as a symmetric matrix (TD = SY). A diagonal matrix (TD = DI) could have presented a simpler model where all covariances are zero. However, this example illustrates how some of these parameters can be estimated.

Error in Variables – Analysis of Covariance Structure

173

LK, on the next line, stands for the label of the ksi constructs, although there is only one of them in this example. That label “FactorOne” follows on the next line. The following line starting with FR is the list of the parameters that are estimated where LX stands for lambda x and TD for theta delta. Each are followed by the row and column of the corresponding matrix, as defined in the model specification in Equations (8.26) and (8.30). The line “Path Diagram” indicates that a graphical representation of the model is requested. The last line of the input file describes the output (OU) requested. SE means standard errors, TV their t-values and MI the modification indices. The LISREL8 output of such a model is given in Figure 8.3. In the output, as shown in Figure 8.3, after listing the instruction commands described earlier

174

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

175

176

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

177

178

Statistical Analysis of Management Data

according to the model specified in the input file shown in Figure 8.2, the observed covariance matrix (in this case a correlation matrix) to be modeled is printed. The “Parameter Specifications” section indicates the list and number of parameters to be estimated, with a detail of all the matrices containing the parameters. The value zero indicates that the corresponding parameter is fixed and is not to be estimated. Unless specified otherwise, the default value of these fixed parameters is set to zero. The number of iterations shows the number which was necessary to obtain convergence and the parameter estimates follow. Below each parameter estimate value, its standard error is shown in parentheses and the t-value below it. Then follow the Goodness of Fit Statistics among which those described earlier can be found. This example run in Figure 8.3 shows that the single factor model represents well the observed correlation matrix since the chi squared is not statistically significant and the GFI is high with a value of 0.98. The modification indices are reasonably small, which indicates that freeing additional parameters would not lead to a big gain in fit. The diagram of such a confirmatory factor analytic model is shown in Figure 8.4 below. 8.3.2

Example of Model to Test Discriminant Validity Between Two Constructs

The following example is typical of an analysis where the goal is to assess the validity of a construct. A construct must be different from other constructs (discriminant validity) but are nevertheless mutually conceptually related (convergent validity). The discriminant validity of the constructs is ascertained by comparing measurement models where the correlations between the construct is estimated with one where the correlation is constrained to be

Error in Variables – Analysis of Covariance Structure

179

one (whereby assuming a single factor structure). The discriminant validity of the constructs is examined for each pair at a time. This procedure, proposed by Bagozzi, Yi and Phillips (1991) indicates that, if the model where the correlation is not equal to one improves significantly the fit, the two constructs are distinct from each other, although they can possibly be significantly correlated. The convergent validity of the constructs is assessed by comparing a measurement model where the correlation between the two constructs is estimated with a model where the correlation is constrained to be equal to zero. A significant improvement in fit indicates that the two constructs are indeed related, which confirms convergence validity. Combining the two tests (that the correlation is different from one and different from zero) demonstrates that the two constructs are different (discriminant validity) although related with a significantly different from zero correlation (convergent validity). Figure 8.5 shows the input file to estimate a two factor model (such analyses are usually performed two factors at a time because the modeling of all the factors at once typically involves problems too big to obtain satisfactory fits). The commands are identical to those described with Figure 8.2, except that now two constructs, “FactorOne” and “FactorTwo”, are specified. The LISREL8 ouput corresponding to this two-factor confirmatory factor structure is shown in Figure 8.6. The description of this output is similar to the one described above involving a single factor. The major difference is the estimate of the correlation between the two factors, which is shown to

180

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

181

182

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

183

184

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

185

186

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

187

188

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

189

190

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

191

192

Statistical Analysis of Management Data

be –0.56 in this particular example. The diagram representing that factor analytic structure is shown in Figure 8.7. Figures 8.8 and 8.9 show respectively the input and output files for a factor analytic structure where a single factor is assumed to be reflected by all the items. The resulting chi squared in Figure 8.9) can be compared with the chi-squared resulting from a model with a correlation between the two factors in Figure 8.6). The difference (126.75 – 54.78) has one degree of freedom and its significance indicates that there are indeed two different constructs (factors), i.e., demonstrating the discriminant validity of the constructs. Next, in order to assess the convergent validity, one needs to compare the fit of a model with zero correlation between the factors with a model where the factors are correlated (as in Figure 8.6). The input and output files for a model with independent factors (zero correlation) are shown respectively in Figures 8.10 and 8.11. The independent factor model has a chi squared

Error in Variables – Analysis of Covariance Structure

193

of 84.34 (Figure 8.11), which when compared with the chi squared of the model estimating a correlation between the two constructs (Figure 8.6) shows a chi squared difference of 29.56. This difference being significant (with one degree of freedom at the 0.05 level), this indicates that the constructs are not independent, i.e., showing convergent validity of the two constructs.

194

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

195

196

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

197

198

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

199

200

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

201

202

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

203

204

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

205

206

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

207

208

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

209

210

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

211

212

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

213

214

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

215

216

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

217

218

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

219

220

Statistical Analysis of Management Data

Instead of defining the variances of the unobserved constructs to unity, the result would have been the same if one lambda for each construct had been fixed to one but the variances of these constructs had been estimated. This is illustrated with the input which would be needed for running this model with AMOS (although it can been done easily with LISREL8 following the principles described above, this example uses AMOS to introduce its commands. The input of the corresponding two factor confirmatory factor model with AMOS is shown below: In AMOS, such as shown in Figure 8.12, each equation for the measurement model (as well as for structural relationships) can be represented with a variable on the left side of an equation and a linear combination of other variables on the right hand side. These equations correspond to the equations as specified by Equations (8.22), (8.25) and (8.26). Inserting “(1)” before a variable on the right hand side indicates that the coefficient is fixed to that

Error in Variables – Analysis of Covariance Structure

221

value and that the corresponding parameters will not be estimated. The program recognizes automatically which variables are observed and which ones are unobserved. Correlations are indicated by “variable1 <> variable2” , where variable1 and variable2 are the labels of observed variables or of hypothetical constructs. The output provides similar information as available in LISREL8.

8.3.3

Example of Structural Model with Measurement Models

The examples above were concerned exclusively with a measurement model or confirmatory factor analysis. As introduced earlier in this chapter, this is only one component of Analysis of Covariance Structures. The full model contains structural relationships among the unobserved constructs that need to be estimated. An example is provided below, where two characteristics of innovations (the extent to which an innovation is radical and the extent to which it is competence enhancing) are hypothesized to affect two constructs, one being changes in the management of the organization and the other being the success of that organization. Figure 8.13 presents the LISREL8 input file for the first step in the analysis, i.e., the measurement model for all the constructs (including both the exogenous and endogenous constructs, although it would be feasible to estimate a separate measurement model for each). The

222

Statistical Analysis of Management Data

results obtained and shown in Figure 8.15 (and represented graphically in Figure 8.14) are then used as input for step 2, which consists in estimating the structural model parameters with the measurement parameters fixed to the values obtained in step 1. This LISREL8 input file is shown in Figure 8.16.

Error in Variables – Analysis of Covariance Structure

223

224

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

225

226

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

227

228

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

229

230

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

231

232

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

233

234

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

235

236

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

237

The estimation of the model presented in that Figure leads to maximum likelihood structural parameter estimates which take into consideration the fact that the constructs are measured with error. These parameter estimates are shown graphically in Figure 8.17 and the full LISREL8 output is listed in Figure 8.18. These examples are given for illustrative purposes; the model shown here could be improved, as its fit to the data is not as high as would be desirable.

238

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

239

240

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

241

242

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

243

244

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

245

246

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

247

248

Statistical Analysis of Management Data

Error in Variables – Analysis of Covariance Structure

249

250

Statistical Analysis of Management Data

8.4 Assignment Using the SURVEY data, develop a model that specifies structural relationships between unobservable constructs measured with multiple items. Develop a model with multiple equations and verify the identification of the structural model. Estimate first the measurement model corresponding to a

Error in Variables – Analysis of Covariance Structure

251

confirmatory factor analysis (including convergent and discriminant validity) and then estimate the structural model parameters. References Basic Technical Readings Anderson, J. C. and D. W. Gerbing (1988), “Structural Equation Modeling in Practice: A Review and Recommended Two-Step Approach,” Psychological Bulletin, 103, 3, 411–423. Bagozzi, R. P. and Y. Yi (1988), “On the Evaluation of Structural Equation Models,” Journal of the Academy of Marketing Science, 16 (Spring), 74–94. Bagozzi, R. P., Y. Yi and L. W. Phillips (1991), “Assessing Construct Validity in Organizational Research,” Administrative Science Quarterly, 36, 421–458. Baumgartner, H. and C. Homburg (1996), “Applications of Structural Equation Modeling in Marketing and Consumer Research: A Review,” International Journal of Research in Marketing, 13 (April), 139–161. Bearden, W. O., S. Sharma and J. E. Teel (1982), “Sample Size Effects on Chi Square and Other Statistics Used in Evaluating Causal Models,” Journal of Marketing Research, 19 (November), 425–430. Bentler, P. M. (1980), “Multivariate Analysis with Latent Variables: Causal Modeling,” Annual Review of Psychology, 31, 419–456. Bentler, P. M. and D. G. Bonett (1980), “Significance Tests and Goodness of Fit in the Analysis of Covariance Structures,” Psychological Bulletin, 88, 3, 588–606. Gerbin, D. W. and J. C. Anderson (1987), “Improper Solutions in the Analysis of Covariance Structures: Their Interpretability and a Comparison of Alternate Respecifications,” Psychometrika, 52,1, 99–111. Gerbing, D. W. and J. C. Anderson (1988), “An Updated Paradigm for Scale Development Incorporating Unidimensionality and its Assessment,” Journal of Marketing Research, 25, 2, 186–192. Joreskog, K. G. (1973), “A General Method for Estimating a Linear Structural Equation System,” in Goldberger and Duncan, eds., Structural Equation Models in the Social Sciences, NY: Seminar Press, pp. 85, 85–112.

Application Readings Ahearne, M., T. W. Gruen, C. B. Jarvis (1999), “If Looks Could Sell: Moderation and Mediation of the Attractiveness Effect on Salesperson Performance,” International Journal of research in Marketing, 16, 4, 269–284. Anderson, J. C. (1987), “An Approach for Confirmatory Measurement and Structural Equation Modeling of Organizational Properties,” Management Science, 33, 4 (April), 525–541. Anderson, J. C. and J. A. Narus (1990), “A Model of Distributor Firm and Manufacturer Firm Working Partnerships,” Journal of Marketing, 54 (January), 42–58. Capron, L. (1999), “The Long-Term Performance of Horizontal Acquisitions,” Strategic Management Journal, Forthcoming. Cudeck, R. (1989), “Analysis of Correlation Matrices Using Covariance Structure Models,” Psychological Bulletin, 105, 2, 317–327. Gilbert, F. W. and W. E. Warren (1995), “Psychographic Constructs and Demographic Segments”, Psychology and Marketing, 12, 3 (May), 223–237. Kuester, S., C. Homburg and T. S. Robertson (1999), “Retaliatory Behavior to New Product Entry,” Journal of Marketing, 63, 4 (October), 90–106. MacKenzie, S. B., R. J. Lutz and G. E. Belch (1986), “The Role of Attitude Toward the Ad as a Mediator of Advertising Effectiveness: A Test of Competing Explanations,” Journal of Marketing Research, 23, 2 (May), 130–143. Murtha, T. P., S. A. Lenway and R. P. Bagozzi. (1998), “Global Mind-sets and Cognitive Shift in a Complex Multinational Corporation,” Strategic Management Journal, 19, 97–114. Philips, L. W. (1981), “Assessing Measurement Error in Key Informant Reports: A Methodological Note on Organizational Analysis in Marketing,” Journal of Marketing Research, 18, 4 (November), 395–415.

252

Statistical Analysis of Management Data

Philips, L. W., D. R. Chang and R. D. Buzzell (1983), “Product Quality, Cost Position and Business Performance,” Journal of Marketing, 47, 2, 26–43. Reddy, S. K. and P. A. LaBarbera (1985), “Hierarchical models of Attitude,” Multivariate Behavioral Research, 20, 451–471. Stimpert, J. L. and I. M. Duhaime (1997), “In the Eyes of the Beholder: Conceptualizations of the Relatedness Held by the Managers of Large Diversified Firms,” Strategic Management Journal 18, 2, 111–125. Titman, S. and R. Wessels (1988), “The Determinants of Capital Structure Choice,” The Journal of Finance, XLIII, 1 (March), 1–19. Trieschmann, J. S., A. R. Dennis, G. B. Northcraft and A. W. Niemi, Jr. (2000), “Serving Multiple Constituencies in Business Schools: M.B.A. Program Versus Research Performance,” Academy of Management Journal, 43, 6, 1130–1141. Vanden, A. P. (1989), “Comment on ‘An Investigation of the Structure of Expectancy-Value Attitude and its Implications’,” International Journal of Research in Marketing, 6, 85–87. Venkatraman, N. and V. Ramanujam (1987), “Planning System Success: A Conceptualization and an Operational Model,” Management Science 33, 6 (June), 687–705. Walters, R. G. and S. B. MacKenzie (1988), “A Structural Equations Analysis of the Impact of Price Promotions on Store Performance,” Journal of Marketing Research, 25 (February), 51–63. Yi, Y. (1989), “An Investigation of the Structure of Expectancy-Value Attitude and its Implications,” International Journal of Research in Marketing, 6, 71–83. Yi, Y. (1989), “Rejoinder to ’An Investigation of the Structure of Expectancy-Value Attitude and its Implications’,” International Journal of Research in Marketing, 6, 89–94.

9. Analysis of Similarity and Preference Data

Similarity data in management research are typically collected in order to understand the underlying dimensions determining perceptions of stimuli such as brands or companies. One advantage of such data is that it is cognitively easier for respondents to provide subjective assessments of the similarity between objects than to rate these objects on a number of attributes. Furthermore, when asking respondents to rate objects on attributes, the selection of the attributes proposed may influence the results while, in fact, it is not clear that these attributes are the relevant ones. In Multidimensional Scaling, the methodology allows one to infer the structure of perceptions by enabling the researcher to make inferences regarding the number of dimensions necessary to fit the similarity data. In this chapter, we first describe the type of data collected to perform multidimensional scaling and we then present metric and non-metric methods of multidimensional scaling. Multidimensional Scaling explains the similarity of objects. We then turn to the analysis of preference data, where the objective is to model and explain preferences for objects. These explanations are based on the underlying dimensions of preferences that are discovered through the methodology. 9.1

Proximity Matrices

The input data for multidimensional scaling correspond to proximity or distance measures. Several possibilities exist, especially metric versus nonmetric and conditional versus unconditional. 9.1.1

Metric versus Non-metric Data

The data which serve as input to similarity analysis can be metric or nonmetric. Metric measures of proximity are ratio scales where zero indicates perfect similarity of two objects. The scale measures the extent to which the objects differ from each other. This measure of dissimilarity between objects is used as input to the method which will consist in finding the underlying dimensions that discriminate between the objects to reproduce the dissimilarities (or similarities) between objects. In effect, these measures are distance measures (dissimilarity) or proximity measures (similarity), and the objective is to produce the map underlying the distances between the objects. Non-metric data reflects also these proximity measures; however, only information about the rank order of the distances is available. As discussed

254

Statistical Analysis of Management Data

in Chapter 7, special care must be taken with such data because most standard statistics such as means, standard deviations and correlations are inappropriate. 9.1.2

Unconditional versus Conditional Data

With unconditional data, all entries in the rows and columns are comparable, i.e., each stimulus is ranked relative to all other stimuli in the matrix (a number from 1 to n(n – 1)/2 for non-metric data). If only the entries within a particular row are comparable, i.e., each of the n column stimuli is ranked relative to one row stimulus (a number from 1 to n for non-metric data), the data are said to be conditional. In this case, the data matrix consists of n – 1 objects ranked in terms of similarity relative to the row stimulus. Even though it is less cognitively complex for respondents to provide conditional data, unconditional data are frequent. 9.1.3

Derived Measures of Proximity

It should be noted that it may be possible to derive distance measures from data of the evaluation of stimuli on attributes. However, it is not clear what attributes should be used and why some other relevant ones may be missing. Furthermore, if the objective is to assess the underlying dimensions behind these attributes, multidimensional scaling will use the computed proximities as input and will ignore some of the information contained in the original attribute level information. Consequently, the use of such a procedure will lose information relative to, for example, principal component analysis. It therefore appears more effective to reserve multidimensional scaling for direct measures of similarity rather than similarity measures derived from attribute level data. 9.1.4

Alternative Proximity Matrices

Apart from these two broad categories of proximity data, the matrix can take several specific forms. 9.1.4.1

Symmetric (half) matrix – missing diag. (=0)

When dealing with distance measures, it is clear that the distance between objects A and B is the same as the distance between objects B and A. Therefore, when concerned with pure distance or proximity data, the full data are contained in half of the matrix, where the rows and the columns denote the objects and the cells represent the distance between these two objects. This matrix is symmetric. Furthermore, the diagonal represents the distance

Analysis of Similarity and Preference Data

255

between an object and itself, and, consequently, the elements of the diagonal are zeros (often they are not even included in the input). 9.1.4.2

Nonsymmetric matrix – missing diag. (=0)

In some cases, the matrix may not be symmetric. This is the case for confusion data, which consists of having each cell represent the frequency with which object i is matched with object j (for example with morse codes, the percentage of times that a code of a particular letter is understood to be some other letter) or one minus that percentage. The greater the confusion, the greater the similarity between the two objects. 9.1.4.3

Nonsymmetric matrix – diag. present

In the case of confusion data, the diagonal may not be zeros because a particular stimulus (e.g., a letter) may not be recognized all the time.

9.2

Problem Definition

In defining the problem, we consider non-metric dissimilarity measures among n stimuli. This is the basic problem such as defined for the KYST algorithm. Let the table or matrix of dissimilarity (input data) be represented by

where is symmetric and the diagonal cells are zero for all j’s). Although we do not know the dimensions of perceptions underlying these distance measures, let us assume there are r such dimensions and that the stimuli are rated on these dimensions. Let be the vector of coordinates of object j in the r-dimensional space. If, indeed, we knew these values

and

r, then we would be able to compute the Euclidean distance between each pair of objects j and k:

The problem is then defined as finding pairs are closest to the actual dissimilarities

such that the

for all

256 9.2.1

Statistical Analysis of Management Data Objective Function

Because the input data about the dissimilarities are not metric, the basic concept used here is to transform the rank-ordered dissimilarities through a monotonic function:

To reproduce the original dissimilarity data, the calculated Euclidean distance should lead to a rank order of these similarities as close as possible to the original or, equivalently, there should be a monotonic transformation of the rank-ordered dissimilarities that are as similar as possible to the computed distances. The differences between the monotonic transformation of the rankordered dissimilarities and the calculated dissimilarities are the error in the fit for each pair i, j:

which, for all the pairs, gives the function to minimize:

This quantity above is divided by a scaling factor, usually order to interpret the objective function relative to the distance values:

9.2.2

in

Stress as an Index of Fit

Equation (9.6) provides the basis of the measure or index of fit of the model at the optimal level. This measure is called the stress and is obtained as

Where M = index for each object pair from 1 to MM DIST(M) = computed distances from the solution of DHAT = predicted distances obtained from the monotonic regression of DIST on the rank-ordered dissimilarity data, DBAR = arithmetic average of the values of variable DIST. The denominator enables the comparison across solutions with a different number of dimensions r.

Analysis of Similarity and Preference Data

257

Equation (9.7) can be rewritten as

where

It is clear from Equations (9.7) or (9.8) that a stress of 0 indicates a perfect fit. 9.2.3

Metric

The discussion above used Euclidean distance measures

This is the most commonly used metric. However, it is possible also to use the Minkowski p-metric:

The easiest case to interpret is for p = 1, which represents the city block metric. For p = 2, it is the Euclidean distance. 9.2.4

Minimum Number of Stimuli

A minimum number of data points (distances) is needed to be able to derive a space that can reproduce the distances. This number has been empirically assessed to be between 4 and 6 objects per dimension. Even though the researcher does not know a priori the number of dimensions, that means that a significant number of objects are needed to implement the methodology successfully. However, because the most typical solutions involve two or three dimensions, a dozen to eighteen objects should be sufficient in most cases.

258

9.2.5

Statistical Analysis of Management Data Dimensionality

Because the number of dimensions r is not known a priori, and because the solution for the depends on the number of dimensions, the dimensionality must be inferred from the results obtained for different values of r. Three criteria can be used together: The stress levels under different dimensionality assumptions, the stability of the results, and the interpretability of these solutions. The goodness of fit or stress values can be plotted as a function of the number of dimensions (scree plot) to identify the elbow where adding dimensions produces little marginal gain in stress levels:

The stability of the results is typically assessed by splitting the sample in two and verifying that the results are similar for each subsample. The interpretability of the results concerns the meaning of the dimensions of perception uncovered by the procedure. Although subjective, this is the most critical for the research to be meaningful. 9.2.6

Interpretation of MDS Solution

The interpretation of the dimensions is mostly the fruit of the researcher’s expertise. However, this expertise can benefit from a complementary data analysis when the objects have also been rated on a number of attributes (although this does lengthen considerably the task for the respondents). This analysis consists of property (attribute) fitting procedures. Three possibilities are available: (a) Maximum r procedure This is based on the bivariate correlation coefficient of each attribute with a particular dimension. A high value of the correlation indicates a strong linear relationship between that attribute and the dimension. Consequently, this attribute would provide a significant input in the identification of the dimension. (b) Monotone multiple regressions A combination of attributes can explain the dimension in a non-linear fashion. The provide a measure of the explanatory power.

Analysis of Similarity and Preference Data

259

(c) Property Fitting (PROFIT) This analysis provides for the possibility ofnon-monotonous relationships. The objective is to obtain a fit so that the stimulus projections are correlated with the scale. 9.2.7

The KYST Algorithm

Finding a solution, as described above, involves finding an initial configuration from which to start an iterative process and then determining the process by which to move from an iteration to the next. Step 1: Finding Initial Configuration Assume that the coordinates are centered at the origin (the means are zero). Let the n objects be identified by their coordinates in the p-dimensional space:

The principal component decomposition of can provide the initial configuration with r eigenvectors or orthogonal dimensions. Step 2: Configuration Improvement In this step, the gradient of the stress provides the direction in which the solution should be changed to improve its value. For that purpose, the disparities between the actual dissimilarities and the predicted dissimilarities computed from the current iteration solution are calculated and the stress S is computed according to the equations above. The gradient is computed from the changes in the stress from one iteration to the next relative to the changes in the coordinate values from

260

Statistical Analysis of Management Data

the prior to the current iteration:

The coordinate values gradient. 9.3

are then modified in the direction of the

Individual Differences in Similarity Judgments

One way to recognize individual differences in perceptions is to allow all m subjects to share a common space, but to permit each individual to weight differently the dimensions of this space (which corresponds to stretching and shrinking of the axes). This assumption is reflected in the INDSCAL algorithm. Consequently, we denote the matrix of dissimilarities between objects for individual i as:

Each individual has a different weight for each dimension. These weights are represented by the diagonal matrix. Let The problem consists now of finding not only the coordinates of points in the common space but also the weights of each dimension for each individual so as to reproduce as much as possible the original dissimilarities:

Wold’s non-linear iterative least squares procedure is used where, at each iteration, either X or is fixed to the last iteration estimate.

9.4

Analysis of Preference Data

In this section we do not refer any longer to modeling for the purpose of understanding the underlying dimensions of perceptions. Now the objective is to represent preferences for some stimili over others. Preferences follow from two basic models. One model predicts that more of any dimension is always preferred to less. This is the Vector model of preference. Another model assumes that the more the better is true up to a certain point, from which too much is as bad as not enough. This assumption corresponds to the Ideal Point model of preference.

Analysis of Similarity and Preference Data 9.4.1

261

Vector Model of preferences

MDPREF is a model which derives the space where stimuli are represented in terms of preferences, as well as the individual differences in preference. Individuals are represented in a preference space by different vectors. Each vector is defined so that the projections of the brands/stimuli on this vector correspond to this individual’s preferences such that the more the projection falls in the direction of the vector, the more preferred the stimulus. The stimuli are represented in the space by points such that the projections on the individual vector correspond the closest possible to the stated preferences. In MDPREF, both the individual vectors and the stimuli points are inferred simultaneously from the preference data. 9.4.2

Ideal Point Model of Preference

PREFMAP differs in two major ways from MDPREF. First, while the individual vectors of preferences and the stimuli points are derived simultaneously from the preference data, this is not the case in PREFMAP. In this program, the stimuli configuration is provided externally. This configuration is obtained from the other methods we described above to derive a perceptual map from similarity data. The results of KYST or INSCAL can be used as input in this analysis of preferences. The second difference comes from the possibility of analysing a vector model of preference as well as ideal point models. Indeed, PREFMAP offers two models of preferences. The vector model is similar to the model described above in the context of MDPREF. However, the difference is, as discussed above, due to the fact that the stimuli points are externally supplied. The interpretation of the individual vectors is similar to what is described above. However, the interpretation of the stimuli configuration is more easily done, as the configuration corresponds to perceptions and not preferences. The joint space for representing perceptions and preferences facilitates also the interpretation of the individual vectors since the dimensions are those derived from the perceptual analysis. The ideal point model of preferences is such that preferences for an individual are also represented as a point in the perceptual space. The preferences for stimuli are such that the most preferred are the stimuli that are the closest in that space to the point representing the individual ideal preference. The further away the stimuli are from the ideal point, the less preferred they are. PREFMAP derives the ideal points for each individual that best represent his/her preference. It should be noted that the vector model is a particular case of the ideal point model where the ideal point is located at infinity.

9.5

Examples

Examples of the various algorithms described above are now given using the PC-MDS software.

262 9.5.1

Statistical Analysis of Management Data Example of KYST

Rank-ordered measures of dissimilarity between brands are the major input of KYST. The example input file is shown in Figure 9.1. The first line of the input file contains three numbers. The first number is the number of stimuli (here, 10 brands). The second number and the third number are for the number of replications and the number of groups (usually 1 each). The second line is the format (Fortran style) in which the data will be read. The data matrix is then shown with 9 rows and 9 columns of the bottom half of a symmetric matrix without the diagonal (assumed to be zeros). Finally, the stimuli (here, brands) labels are written on separate lines. The output of KYST with this particular problem is shown in Figure 9.2. A two-dimensional solution was requested during the interactive dialog while running the software by indicating a minimum and a maximum number of dimensions of 2. The output shows the results by providing the stress obtained from that solution (a stress value of 0.266) and the coordinates in that twodimensional space for the ten brands. The Shepard diagram represents the plot of the pairs of brands with the actual dissimilarity data on the y axis and the computed distances (before and after transformation through monotone regression). This shows how well the model replicates each of the pairs of

Analysis of Similarity and Preference Data

263

264

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

265

266

Statistical Analysis of Management Data

stimuli. The plot of the brands in the two-dimensional space is shown, where the brands are numbered in the order of the input. The interpretation can be inferred from the knowledge about the brands according to the attributes that appear to discriminate these brands along the two dimensions found (here, an economy and a performance dimension). An example of PROFIT analysis to help interpret the meaning of the dimensions is shown below. 9.5.2

Example of INDSCAL

In INDSCAL, the data for several individuals are analyzed. The input file of an example is shown in Figure 9.3.

Analysis of Similarity and Preference Data

267

268

Statistical Analysis of Management Data

The first line of the input file contains the following information: Number of ways of the data (3-way data) Maximum number of dimensions (2 in this example) Minimum number of dimensions (2 in this example) Type of input data (2 means lower-half dissimilarity matrix with no diagonal; other possibilities include a value of 1 for a lower-half similarity matrix without diagonal) Maximum number of iterations (25 were defined in this example). The remaining codes on this first line correspond to more advanced options. The second line contains a number for each way. The first one is the number of subjects and the other two give the number of stimuli. The third line shows the format (Fortran-style) in which the data will be inputed. The dissimilarity data are then shown for each individual (it is good practice to show first the subject number, although, as indicated by the format statement, this number is not read in). Finally, the objects labels (brand names) are listed, one per line. The results of INDSCAL are shown in Figure 9.4.

Analysis of Similarity and Preference Data

269

270

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

271

272

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

273

274

Statistical Analysis of Management Data

The output, under the title “history of computation,” shows the fit measure at each iteration. Because INDSCAL is a metric model, the fit measure is the correlation between the input dissimilarity data and the predicted dissimilarity from the model parameter values at that iteration. The value of 0.999 obtained in the example is excellent.

Analysis of Similarity and Preference Data

275

Under the title “Normalized A Matrices” matrix 1 lists the individual weights for each of the 4 individuals. Matrix 2 lists the coordinates of the objects in the common object space. The individual weights shown in Matrix 1 are plotted along the two dimensions in the first plot. Plot No. 2 represents the brands corresponding to the coordinates listed in matrix 2.

9.5.3

Example of PROFIT (Property Fitting) Analysis

In the example below, we use the configuration (coordinates) obtained from the KYST analysis described earlier in section 9.5.1. (It is possible to use the output configuration of other models such as INSCAL). The relationships of the two dimensions corresponding to these perceptions of the ten brands with five characteristics of the brands (i.e., weight, design, volume, maximum frequency and power) are analysed in this run of PROFIT. Therefore, the ratings of these brands on these characteristics are matched as well as possible with the ratings obtained from the KYST configuration. Each characteristic is represented in the perceptual space by a vector so that the fit with the perceptions of the brands is maximized. For rating data on the properties (brand characteristics), the correlation between these ratings and the projection of the brand perceptions on that vector is maximized. The input file shown in Figure 9.5 provides the information necessary to run the program. The first line of input indicates the basic parameters of the problem. The first number (1 in Figure 9.5) indicates that a linear relationship between properties and perceptions will be evaluated. The second number (10 in Figure 9.5) indicates the number of stimuli (brands). The third number (2 in Figure 9.5) shows the number of dimensions in the perceptual space used as input. The fourth number (5 in Figure 9.5) is the number of properties to be analyzed. The other numbers correspond to more advanced options. The second line is the Fortran-style format in which the data for the stimuli (brands) coordinates are read. Then follow the perception coordinates, one line for each stimulus (brand). In this example, the stimulus number (1 to 10) is shown to better visualize the input, but this information is not read by the program, as the format above indicates that the first two columns are skipped (“2X”). After the perceptual coordinates, the data on the properties are shown. First, the format in which the data are to be read is indicated in the ususal Fortran-style format. Then, for each of the properties, the label of the property is shown on a line and on a separate line the values of the property on all the ten stimuli are shown. The first number indicates the property number but is not used, as shown by the format of the input which skips the first two columns of data (“2X”). Finally, the last ten lines correspond to the labels of the ten stimuli, in this case the names of the brands.

276

Statistical Analysis of Management Data

Figure 9.6 shows the output of the PROFIT analysis. First, for each property, the correlations between the the original and the fitted vectors are shown, followed by the corresponding plot of the stimuli. The last graph shows the perceptions of the stimuli (the ten brands) numbered from 1 to 9, plus the letter A to represent the tenth brand. The points labeled B to F represent the end points of the property vectors that maximize the correlation with the projections of the brands on this vector with the original property values. The vectors have been added in Figure 9.6 and do not appear on the original computer output. B represents the weight property, C represents the design, D the volume, E the maximum frequency, and F the power of the brands.

Analysis of Similarity and Preference Data

277

278

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

279

280

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

281

282

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

283

284

Statistical Analysis of Management Data

Fig. 9.6

Output example for INDSCAL (examp9-3.out).

This plot indicates that the Y dimension (dimension 2) is closely related to Weight and Power and also, although not as strongly, to Design (the higher the values of the properties, the lower the perceptions on that dimension). The X dimension (dimension 1) reflects more the volume, which appears to be negatively correlated with the maximum frequency. Therefore, generally, the higher the perceptual value on dimension 1, the higher the volume but the lower the maximum frequency. It should be noted that these can be used only to help the interpretation of the dimensions. However, the dimensions and the properties do not coincide perfectly. For example, although the vectors B and F are particularly close to axis Y, Axis X is not very close to either vector

Analysis of Similarity and Preference Data

285

D or E. Consequently, the property fitting analysis will not be as useful to interpret the X axis as it will be for the Y axis. 9.5.4

Example of MDPREF

The first row in the input file shown in Figure 9.7 defines: the number of rows in the data matrix, or number of subjects (there are 5 subjects in this example); the number of columns in the data matrix, or number of stimuli (there are 10 brands shown in Figure 9.7); number of dimensions (2 in this example); number of dimensions to be plotted (2); a code to normalize by subtracting the row mean (= 1) or to normalize and divide by the standard deviation (= 2); a dummy code to normalize subject vectors (= 1; 0 otherwise). The second line defines the format in which the preference data are read, followed by the data themselves. The first number of each row is the subject number, which is not read by the program, as indicated by the format statement starting with 2X. For each row (subject), the ten numbers indicate the values given by the subject to each of the 10 brands.

286

Statistical Analysis of Management Data

The following lines are used for the labels of the subjects and then of the stimuli. The first graph in the output file maps the subject vectors starting at the origin with the end point at the location of the number corresponding to the subject. The second graph maps the stimuli according to their preferences, while the third graph shows both the subject vectors and the stimuli points at the same time. Given that the sole input concerns preferences, this plot of the brands should be carefully interpreted, as it does not correspond to perceptual data but is only derived from preferences. On the graphs shown in Figure 9.8, the vectors have been added to the original output. The projections of the stimuli on a particular subject vector indicate the preferences of that individual subject. For example, subject 1 (indicated by the letter B on Figure 9.8) has a preference for brands 3 (SEMI) and 7 (SONO). Subject 5 (letter F on Figure 9.8) prefers brand 10 (SUSI; indicated by the letter A) and then brands 1 (SAMA) and 6 (SIRO), both confounded on the map and represented by the symbol “#”. The least preferred brands for this subject are brands 2 (SALT), 8 (SOLD) and 9 (SULI), these last two brands being confounded on the map and represented by the # sign in the lower right quadrant.

Analysis of Similarity and Preference Data

287

288

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

289

290

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

291

292 9.5.5

Statistical Analysis of Management Data Example of PREFMAP

In the example provided in Figure 9.9, the external source of the perceptual space configuration has been taken from the INDSCAL run. The first line of input in that file allows the user to define the various parameters concerning the data and the analysis to be done: the number of stimuli (here 10 brands) the number of dimensions of the externally supplied perceptual space (here 2) the number of subjects for which preferences are being modeled (here 5)

Analysis of Similarity and Preference Data

293

a code to indicate that the higher the score of a brand in the data, the higher the preference for that brand (code = 1) or that the higher the score, the lower the preference (code = 0); in the example, preferences are decreasing with the ratings and, therefore, a code 0 has been entered. These numbers are followed by additional codes corresponding to advanced setting options. The second line of input gives the format in which the coordinates in the perceptual space will be read. Then follow these coordinates for the ten stimuli/brands. Note that, given the format provided, the stimulus number (the first number on each of the line for the coordinates) is not read by the program. Then follows the format in which the preference data will be read. These preference data correspond to the ones described for the input of MDPREF. Therefore, the preference ratings of the ten brands are shown for each of the four subjects studied. Finally, the stimuli labels (brand names) are indicated. The results are shown in Figure 9.10. Phase 1 corresponds to the general unfolding model where the axes may be rotated differently for each subject and where each subject can weight each axis differently. Although it makes the visualization difficult, due to the different rotation of the axes, this is the most flexible model. It should be noted that there is one more point for subjects than there are subjects. This last point corresponds to the average preference (average ratings) across all the subjects. Phase 2 corresponds to the weighted unfolding model wherein all subjects share the same configuration without rotation but each subject is allowed to weight each dimension differently. The preferences of each subject are shown by his/her ideal point in that common perceptual space. In Phase 3, each subject uses the same perceptual space configuration with no axis rotation and no differential weighting of the dimensions. Finally, Phase 4 corresponds to the vector model of preferences, similarly to MDPREF, except for the fact that the perceptual configuration is externally provided. Here is an example from the INDSCAL analysis. The plot resulting from the analysis of Phase 3 provides the ideal points of the five subjects, as well as that of the average subject. This plot shows that subject 4 (represented by the letter D) prefers brands 2 (SALT), 3 (SEMI), 8 (SOLD) or 9 (SULI) best (the closests to his/her ideal brand). This fits the preference data used as input, where these brands have a low score value (most preferred). For the vector model of preferences, the last graph shows the end points of the individual vectors. The vectors drawn on Figure 9.10 have been added to the original output, showing the differences in preferences across individuals according to the projections of the stimuli on their respective vectors. For example, the projections of the brands on the vectors of subjects 2 (C) and 5 (F), indicate that brands 1 (SAMA), 6 (SIRO) and 10 (SUSI; indicated by the letter A on the plot) are the prefered ones. These correspond indeed to the lowest scores (most preferred) in the input data.

294

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

295

296

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

297

298

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

299

300

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

301

302

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

303

304

Statistical Analysis of Management Data

Analysis of Similarity and Preference Data

305

306

Statistical Analysis of Management Data

9.6

Assignment

Collect proximity data about a set of brands of your choice and determine the dimensions used in the perception of these brands. Gather data about characteristics of these brands to help you interpret the underlying perceptual dimensions. For these same brands, obtain preferences of the respondents in order to develop a map of subject preferences and stimuli.

Analysis of Similarity and Preference Data

307

References Basic Technical Readings Carroll, J. D. and P. Arabie (1980), “Multidimensional Scaling,” Annual Review of Psychology, 31: 607–649. Kruskal, J. B. and M. Wish (1978), Multidimensional Scaling, Beverly Hills, CA: Sage Publications. Shepard, R. N. (1980), “Multidimensional Scaling, Tree-Fitting, and Clustering,” Science, 210, 24 (October), 390–398. Ward, J. (1963), “Hierarchical Grouping to Optimize an Objective Function,” Journal of the American Statistical Association, 58, 236–244.

Application Readings Bijmolt, T. H. A. and M. Wedel (1999), “A Comparison of Multidimensional Scaling Methods for Perceptual Mapping,” Journal of Marketing Research, 36, May, 277–285. Cooper, L. G. (1983), “A Review of Multidimensional Scaling in Marketing Research,” Applied Psychological Measurement, 7, 4 (Fall). DeSarbo, W. S. and G. De Soete (1984), “On the Use of Hierarchical Clustering for the Analysis of Nonsymmetric Proximities,” Journal of Consumer Research, June, 601–610. DeSarbo, W. S., M. R. Young and A. Rangaswamy (1997), “A Parametric Multidimensional Unfolding Procedure for Incomplete Nonmetric Preference/Choice Set Data in Marketing Research,” Journal of Marketing Research, 34, 4 (November), 499–516. Green, P. E: (1975), “Marketing Applications of MDS: Assessment and Outlook,” Journal of Marketing, 39 (January), 24–31. Green, P. E. and F. J. Carmone (1989), “Multidimensional Scaling: An Introduction and Comparison of Nonmetric Unfolding Techniques,” Journal of Marketing Research, 6 (August), 330–341. Helsen, K. and P. E. Green (1991), “A Computational Study of Replicated Clustering With an Application to Market Segmentation,” Decision Sciences, 22, 1124–1141. Johnson, R. M. (1971), “Market Segmentation: A Strategic Management Tool,” Journal of Marketing Research, (February), 13–18. Neidell, L. A. (1969), “The Use of Nonmetric Multidimensional Scaling in Marketing Analysis,” Journal of Marketing, 33 (October), 37–13. Sexton, D. E. Jr. (1974), “A Cluster Analytic Approach to Market Response Functions,” Journal of Marketing Research, February, 109–. Srivatsava, R. K., R. P. Leone and A. D. Shocker (1981), “Market Structure Analysis: Hierarchical Clustering of Products based on Substitution in Use,” Journal of Marketing, Summer, 38–48.

This page intentionally left blank

Appendices Appendix A: Rules in Matrix Algebra Vector and Matrix Differentiation

Kronecker Products

310

Statistical Analysis of Management Data

APPENDIX B: Statistical Tables Cumulative Normal Distribution

Appendices

311

312

Statistical Analysis of Management Data

Appendices

F Distribution

313

314

Statistical Analysis of Management Data

Appendix C: Description of Data Sets

The data sets described below can be downloaded from the web at http://www.insead.edu/~gatignon. Three different kinds of information, which correspond to typically available data about markets, are provided for analysis: industry, panel and survey data. In addition, scanner data is provided for a product category in the form typically available in practice. The industry dataset includes aggregate product and market data for all of the brands sold in each time period. This type of information is often provided by market research services, trade and business publications, and trade associations, to all of the firms competing in an industry. The other two datasets contain information collected from a sample of consumers rather than from the entire population. The first, panel data, is gathered from a group of consumers who have agreed to periodically record their brand perceptions, preferences, and purchase behavior. This information is often purchased by advertisers from syndicated research services and is useful for tracking changes in consumer behavior over time. The second, survey data, is collected by questionnaire or personal interview from a large group of consumers. Surveys are often conducted by advertising agencies (such as DDB Needham Worldwide, N. W. Ayer, and others), survey research companies, and by the advertisers themselves. These surveys typically measure a broad

Appendices

315

range of consumer characteristics, including attitudes, interests, values, and lifestyles. This information is especially useful for selecting target audiences and designing creative appeals. The MARKSTRAT® market simulation program was used to create the industry and panel datasets. The survey dataset was developed separately to conform to this environment. We first describe the MARKSTRAT® environment and the characteristics of the industry. We then present the three types of data provided with this book and discuss the contents of each dataset.

The MARKSTRAT® Environment To understand the industry in which competing firms operate, the reader must be familiar with two general dimensions of the MARKSTRAT® environment: (1) the structure of the industry in terms of the products, competition, and market characteristics, and (2) the marketing decisions that each firm can make over time. The discussion which follows concentrates on those aspects which are most relevant to advertising planning decisions. Competition and market structure In the MARKSTRAT® environment, five firms compete in a single market with a number of brands. Each firm starts out with a set of brands and has the ability to initiate research and development (R&D) projects to create new brands. If an R&D project is successful, then the sponsoring firm has the option of bringing the new product to market. All new products are introduced with new brand names. Product Characteristics. The generic products in this industry are consumer durable goods comparable to electronic entertainment products. They are called Sonites. Because these products are durable, each customer win usually purchase only one item over a long period of time. Consequently, there are no issues of repeat purchase, brand loyalty, or brand switching in this market. The products are characterized by five physical attributes: (1) weight (in kilograms), (2) design (measured on a relative scale), (3) volume (in cubic decimeters), (4) maximum frequency (in kilohertz), and (5) power (in watts). Not all attributes are equally important to consumers. Different segments have different preferences for these product characteristics, although the preferences are expressed in terms of brand image rather than purely physical characteristics. Consumers’ brand evaluations are a function of their perceptions of the brands on three general dimensions, roughly corresponding to three of the five physical characteristics listed above. The first and most important characteristic is the perceived price of the product. Next, people consider

316

Statistical Analysis of Management Data

the product’s power (wattage). Finally, consumers evaluate the product’s design (aesthetic value). Although less important than the other dimensions, the product’s design helps consumers to differentiate between the various competing brands. The design attribute is measured on a scale from 1 to 10 by expert judges. To form an overall evaluation of each brand, consumers compare the brand’s performance on each dimension with their preferences for a certain “ideal level” on each of these dimensions. Because of the durability of the Sonite product and the importance of the purchase, the consumer decision process tends to follow a “high involvement” hierarchy. Measures of brand awareness, perceptions, preferences, and purchase intentions are, therefore, particularly relevant to the advertising decisions. Consumer Segments. The consumer market for Sonites can be decomposed into five segments with distinguishable preferences. Segment 1 consists primarily of the “buffs,” or experts in the product category. They are innovators and have high standards and requirements in terms of the technical quality of the product. Segment 2 is composed of “singles” who are relatively knowledgeable about the product but somewhat price sensitive. “Professionals” are found mostly in segment 3. They are demanding in terms of product quality and are willing to pay a premium price for that quality. “High earners” constitute segment 4. These individuals are also relatively price insensitive. However, they are not as educated as the professionals, and are not particularly knowledgeable about the product category. They buy the product mostly to enhance their social status. The fifth and last segment covers all consumers who cannot be grouped with any of the other four segments. They have used the product less than consumers in other segments and are considered to be late adopters of this product category. Given that this group is defined as a residual, it is very difficult to characterize the members in terms of demographics or lifestyle. Although the preferences of the five consumer segments may change over time, the composition of each segment does not. Consequently, the survey data collected in the eighth time period (to be described) also describes consumers during the previous seven periods. Distribution Structure. Sonites are sold through three different channels of distribution. Each channel carries all brands of Sonites, but the potential number of distributors and the characteristics of each channel are different. Channel 1 is made up of specialty retail stores. These stores provide specialized services to customers, and the bulk of their sales come from Sonites. There are 3,000 such outlets. Electric appliance stores are channel number 2. The 35,000 appliance stores carry Sonite products only as an addition to their main lines of electric appliances. Channel 3 is the 4,000 department stores that exist in the MARKSTRAT® world. Department stores sell a broad range of products, including clothing, furniture, housewares, and appliances. The three channels differ in terms of the proportion of the product that they sell and the types of clientele that they attract.

Appendices

317

Marketing mix decisions A product’s marketing mix reflects the marketing strategy for the brand. A brand’s attributes will influence how the brand is positioned and to whom it is marketed. Its price will affect the advertising budget and the brand image. Its distribution will determine where the brand is advertised, and so on. In this section we review the four main marketing mix variables, price, sales force, advertising, and product, that characterize brands in the MARKSTRAT® environment. Prices. Each Sonite brand has a recommended retail price. These prices are generally accepted by the distribution channels and are passed on to consumers. As indicated earlier, different consumer segments are more or less sensitive to price differences across brands. A segment’s price sensitivity or “elasticity” also depends on the selection of products offered to that segment and on the other marketing mix variables. Sales Force. The two most important aspects of a firm’s sales force are its size and its assignment to the three channels of distribution. Each salesperson carries the entire line of brands produced by his or her company. When a firm changes the number of salespeople it assigns to a particular channel, this is likely to affect the availability or distribution coverage of the firm’s brands. Advertising. Each brand of Sonite is advertised individually. Firms in this industry do not practice umbrella or generic (product category) advertising. However, advertising of specific brands can increase the total market demand for Sonites or affect Sonite demand in one or more segments. Advertising can serve a number of communication purposes. It can be used to increase top-of-mind brand awareness and inform consumers about a brand’s characteristics. Research has revealed that advertising expenditures are strongly positively related to brand awareness. Advertising can also have a substantial persuasive effect on consumers. Advertising can be used to position or reposition a brand so that the brand’s image is more closely aligned with consumers’ needs. In addition, it is clear that advertising plays an important competitive role. One cannot consider a brand’s advertising in isolation. Instead, the relative advertising weight or “share of voice” is a better predictor of consumers’ purchase behavior than absolute advertising expenditures. Share of voice is the ratio of the brand’s advertising expenditures to the total industry spending on advertising. Products. The database reports information on all of the Sonite products that were marketed by firms during an eight-year time period. The names of the brands sold during this period are listed in Table C.1. This table also lists the periods during which each brand was available. The reader should note that some of the brands were introduced after the first time period and /or were discontinued before the last (eighth) period. The brands of Sonites are named to facilitate identification of the marketing firm. The second letter of each brand name is a vowel that corresponds to one

318

Statistical Analysis of Management Data

of the five competing firms. Firm I markets all brands that have an “A” as the second letter of the name, such as SAMA. “E” corresponds to firm 2, “I” to firm 3, “O” to firm 4, and “U” to firm 5. During the eight time periods, each firm has the opportunity to design and market a portfolio of different brands. In response to consumer or market pressures, companies may change the physical characteristics of each brand over time. Information about brands and their attributes is provided in the industry data set, as described below. Survey

A mail survey of a group of 300 consumers was conducted in the eighth (most recent) time period. The survey collected a variety of consumer information including demographic data, psychographics, information on product purchase behavior, decision processes, and media habits. These data are particularly useful for segmentation analysis, which is an important precursor to selecting a target market, generating copy appeals, and media selection. A list of the variables from the questionnaire and the coding scheme for the items are provided in Tables C.2 and C.3 respectively. Indup

The industry dataset provides two types of performance information for each brand and time period: sales figures (in units and dollar sales) and market

Appendices

319

320

Statistical Analysis of Management Data

Appendices

321

322

Statistical Analysis of Management Data

Appendices

323

324

Statistical Analysis of Management Data

share data (based on unit and dollar sales). The dataset also includes information on the values of the marketing mix variables for each competing brand. The data describe each brand’s price, advertising expenditures, sales force size (for each channel of distribution), and physical characteristics (i.e., the four Ps). Finally, the dataset reports the variable cost of each brand at each time period. The reader should note that this cost is not the actual current production cost, as this information is typically not available for each competitive brand. The reported cost figures reflect the basic cost of production that can be estimated for a given first batch of 100,000 units at the period of introduction of the brand. A list of the variables in the industry dataset is given in Table C.4. Panel The panel dataset provides information that, in many ways, complements the data in the industry dataset. Panel data are available at the level of the individual market segment rather than at the total market level. The panel dataset includes information on the size of each segment (in unit sales of Sonites) and the market share for each brand with each segment. The dataset also provides the results of a panel questionnaire with items on advertising communication, brand perceptions, and preferences. Variables include the extent of brand name awareness, segment preferences in terms of the ideal levels of the three most important attributes (price, power, and design), consumers’ brand perceptions on the same three dimensions, and brand purchase intentions. Finally, the dataset reports the shopping habits of each segment in

Appendices

325

the three channels of distribution. A summary of these variables is provided in Table C.5. Scan

SCAN.DAT contains a simulated sample of scanner data, similar to the dataset of refrigerated orange juice dataset used in Fader and Lattin (1993); Fader, Lattin and Little (1992); and Hardie, Johnson and Fader (1992). (See these papers for a full description of this dataset, and SCANNER.SAS for the criteria used to create this subset. The six brands along with their brand id codes, are:

1 2 3 4 5 6

Brand 1 Brand 2 Brand 3 Brand 4 Brand 5 Brand 6

This file is set up for estimation of the standard Guadagni and Little (1983) MNL model of brand choice, including their “loyalty” variable. The value of

326

Statistical Analysis of Management Data

the smoothing constant is set to 0.8; the loyalty variable is initialized using purchase information for weeks 1–52. In this dataset, the number of choice alternatives varies over time (due to shopping at different stores, stock-outs, etc). Rather than having one record per purchase occasion, we have one record per choice alternative. The format of SCAN.DAT is as follows: panelist id week of purchase a dummy variable indicating whether this record is associated with the brand chosen the number of records (brands available) associated with this purchase occasion the brand id of this record regular shelf price for this brand any price reduction for this brand on this purchase occasion (price paid = price – price cut) a dummy variable indicating the presence of a feature ad for this brand the value of the Guadagny and Little loyalty variable for this brand (on this purchase occasion) a brand-specific constant/dummy for brand 1 a brand-specific constant/dummy for brand 2 a brand-specific constant/dummy for brand 3

Appendices

327

a brand-specific constant/dummy for brand 5 a brand-specific constant/dummy for brand 6 The reference brand is, therefore, brand 4, a private label. This dataset was created specifically for analysis using LIMDEP. The file examp6–2.lim contains a sample “program” for reading this dataset into LIMDEP. Note that other estimation packages may require the data in a slightly different format. Minor formatting changes can easily be accomplished using SAS; for major changes it may make sense to modify the SCANNER.PAS program (or write your own version of this program using your language of choice). An executable file “scanner.exe”, is available to create similar data sets using a different constant for building the loyalty variable. Fader, Lattin, and Little (1992) describe a procedure for estimating nonlinear parameters in MNL models using standard MNL estimation routines. The smoothing constant in the G&L loyalty variable is such a nonlinear parameter. The value used in the creation of SCAN.DAT (0.8) is not necessarily optimal. To find the optimal value, you require derivative terms (see Fader, Lattin and Little (1992) for complete details). SCANNER.PAS is the program used to create SCAN.DAT. This program allows you to create a dataset called SCAN.DAT for any value of the smoothing constant (alpha) and gives you the option of including the required derivative terms (which are inserted between the loyalty variable and the brand-specific constant for brand 1; the format is F10.6). The “scanner.exe” program will prompt you for the value to use for the smoothing constant. The original simulated raw data are contained in the files PURCHASES.DAT and STORE.DAT. PURCHASES.DAT contains the simulated purchase history information for 200 households. The format of this file is as follows: panelist id brand id of brand chosen

week of purchase

store id

STORE.DAT contains the store environment information. The format of this file is as follows: week # store id brand id regular price price cut feature dummy Note that a price (and price cut) of 9.99 indicates brand not available that store/week. SCANNER.PUR is the subset of PURCHASES.DAT used in the creation of SCAN.DAT. SCANNER.PUR was created using SCANNER.SAS.

This page intentionally left blank

Index 2SLS see two stage least squares 3SLS see three stage least squares adjusted Goodness of Fit Index (AGFI) 170 advertising 53, 317 expenditures, marketing example 89 Aitken estimator see Generalized Least Squares (GLS) Estimator AMOS 171 two factor confirmatory factor model 220–1 analysis of three variables, data example for 19 types of 29 axis rotation 33–4 Bagozzi, Yi and Phillips (1991) procedure 179 Bartlett test of 92 V of 18, 24 Belgium 21 Bentler and Bonnet (1980) goodness-of-fit index (GFI) 169 Bernoulli process 113 best linear estimator 61–5 Best Linear Unbiased Estimator (BLUE) 62, 81 bivariate normal distribution 9–10 BLUE see Best Linear Unbiased Estimator brand attitude 165 awareness 317 cognitions 166 evaluations of consumers 315 Bretton-Clark 143 business schools, ranking of 154 categorical dependent variables 6, 105 channels of distribution 316 Chi-Square distribution 14 for Independence Model 186, 200, 232 test 120, 168 variate 12 choice modeling 5 classification table 120 coding principle 137

coefficient alpha 31–3 of orthogonal polynomials 142 cognitive responses to advertising 165 commonalities 40 estimating 41 competition and market structure 315–16 complementary data analysis 258 completely restricted residual sum of squares (CRSS) 69–70, 72 completely unrestricted residual sum of squares (CUSS) 69–72 composite measurement with K components, generalization of 33 composite scales 30–3 concentric ellipses 11 conditional logit 119 discrete choice in LIMDEP 118–19 model 117–19 configuration improvement 259 Confirmatory Factor Analysis 29, 39, 42–4 example of 172–8 model, path diagram of 178 Conjoint Analysis 137 using SAS, example 151 Conjoint Designer 143 constitutive indices 38 consumer durable goods 315 information 318 segments 315–17 Consurv 143 contemporaneous covariances, matrix of 80–4 contemporaneously correlated disturbances 79–81 covariance matrix 83, 86 estimation 165–8 test of 92 variances in 40 covariance model 68 covariance structure analysis 6, 29, 159, 221 description of model 163–5 examples of 171 model fit 168–70 covariance structure model, analysis of 171 criterion variable 137,139

330 Cronbach’s alpha 29 cumulative density functions 146 data base INDUP 25–6 PANEL 25–6 data, metric versus non-metric 253 data scan.dat 129 data sets, description of 313 unconditional versus conditional 254 demographics 318 design programs 142–3 developing a composite scale 53 dimensionality 258 discriminant analysis 105–12, 120 classification and fit 109–12 using SAS, example 121–7 discriminant coefficients 127 discriminant criterion 105, 107 discriminant function 108–10 fit measures 110–12 dissimilarity data 256 measure of 253 distribution chi-squared 311–12 coverage 19 cumulative normal 310 outlets 71 structure 316 disturbance-related set of equations 79 dual mediation hypothesis model (DMH) 165 theory 166 Dummy Coding 141–2 dummy variable 141 coding 137 econometric theory elements 55 effect coding 137, 139–42 eigenvalues and eigenvectors 34–7 properties of 36–7 electric appliance stores 316 electronic entertainment products 315 ellipse with centroid 9 endogeneity in system 99 endogenous variables in system 83 England 21 error component model 67 structure 56–7 errors-in-variables, effect of 159

Index Estimated Generalized Least Square (EGLS) Estimator 82, 93, 101, 115 estimators, properties of 61–5 Excel 19 exogenous variables in system 83 Expected Cross-Validation Index (ECVI) 199, 212, 232, 245 Exploratory Factor Analysis 29, 33, 39–43 calculation of eigenvalues and eigenvectors 35 difference from Confirmatory Factor Analysis 38 factor analysis 5–6, 33–43 application examples 44–52 structure 171 factorial design, two by two 137 factors, determining number of 41 Fader, Lattin, and Little (1992) procedure 326 F Distribution 313 firm factors 93 sales force 316 FORTRAN 262 conventions 148 Fortran-style format 275 France 21 Generalized Least Squares (GLS) Estimator 58–60, 62, 81–2, 115–16 general linear models (GLM) procedure 144, 152–3 estimator 62 goodness-of-fit index (GFI) 170, 178 goodness of fit measures 169–70 statistics 178, 185, 199, 212, 231, 245 or stress values 258 grand mean rating 138 graphical representation of full structural model 237 of measurement model for exogenous and endogenous constructs 222 of multiple measures with a confirmatory factor structure 39 of measures 31 Guadagni and Little (1983) MNL model of brand choice 326 heterogeneity of coefficients, issue of 55 hierarchy of effects 104 Hotelling’sT 17

Index Ideal Point model of preference 260–1 identity matrix 92, 99 imperfect measures, impact of 159 IMS 143 independent factor model 192 index of fit of model 256 Individual Differences Scaling (INDSCAL) 261, 268, 275, 284, 292 algorithm 260 analysis 293 example of PC-MDS for 266–74 INDUP.CSV 103, 324 data set 75–6 industry characteristics 93 dataset 324 market segments 75 initial factors, extracting 41 innovations, characteristics of 93 interval scale 3–4 Isodensity contour 9 Iterative Seemingly Unrelated Regression (ITSUR) 93, 99 key learning experiences 6 Kronecker products 81, 309 KYST 261 algorithm 255, 259 analysis 275 example of 262–6 Multidimensional Scaling 263–6 Lagrangian multiplier problem 64 Lawley’s approximation 93 LIMDEP 2, 118, 128–9, 151–6, 326 linear effect coding 140–1 linear model 139 estimation with SAS, examples of 71–4 linear restrictions 65–7 linear statistical model 55–9 LISREL 2, 6,170 Estimates (Maximum Likelihood) 196 LISREL8 for Windows 172–8, 205, 222 in model with single factor 193–205 in model with two factors 179–220 in model with two independent factors 205–19 for full structural model 236–50 for measurement model for exogenous and endogenous constructs 221 LOGIT.EXE 132 logit models of choice 112 LOGIT.PRM 131

331

LOGIT.RES 131–2 logit type model 151 loyalty variable of Guadagny and Little 326 MacKenzie, Lutz and Belch’s Model (1986) of role of attitude towards ad 165–7 mail survey 318–23 main effect model 139 management research, measuring variables in 3 market behavioral responses 25 response function 89, 104 share 19, 25, 71 marketing decision functions 89, 104 mix 25, 316–18 strategy 25,316 MARKSTRAT® market simulation program 314 Environment 314–18 matrix algebra, rules in 309 derivation rule 107 matrix, structure of 91–2 matrix, structure of 92 maximum chance criterion 111 Maximum Likelihood Estimation 43, 59–60 maximum likelihood estimator 60 maximum r procedure 258 MBA program 154 MDPREF see multidimensional analysis of preference data MDS see Multidimensional Scaling means, tests about 12 mean vectors, test of difference between several, K-sample problem 21 between two, one-sample problem 19 measure, definition of a 29 measurement errors, problems associated with 162 measurement model or confirmatory factor analysis 221 parameters 167, 171 results, LISREL8 222–35 measurement theory, notions of 29–33 media habits 321 mental factor analysis 53 Minimum Fit Function Chi-Square 199, 231, 245 Value 245 minimum variance estimator 116 Minkowski p-metric 257

332

Index

model parameters, test of significance of 170 model specification and estimation methods 91 model to test discriminant validity between two constructs 178–221 modification indices 170, 191, 204, 217, 232, 246–8 monotone analysis of variance (MONANOVA) 137, 143–4 using PC–MDS, example 147–51 monotone multiple regressions 258 multidimensional analysis of preference data (MDPREF) example of 261, 285–91 Multidimensional Scaling (MDS) 253 solution, interpretation of 258 via a Generalization of Coombs Unfolding Model 294 multinomial logit analysis using LIMDEP 128–30, 133–5 multinomial logit model 117 analysis using LOGIT.EXE 130–2 multiple independent variables, case with 162 multiple measures, graphical representation of 39 multiple regression with a single dependent variable 55 multivariate analysis of variance 5 multivariate case, generalization to 11 multivariate logit program LOGIT.EXE 130 multivariate normal distribution 9 multivariate test with known 14 with unknown 14–15 nested models 169 nominal scale 3–4 non-linear effect coding 141 non-linear parameters in MNL models 326 non-metric dissimilarity measures 255 Non-Normed Fit Index (NNFI) 232 nonsymmetric matrix diagonal present 255 missing diagonal 255 Normed Fit Index (NFI) 232 Norway 19 objective function 256 ordered model 147 Ordered Probit Analysis using LIMDEP, example 151–6 ordered probit model 144–7 order and rank conditions 89–91,171

ordinal scale 3–4 Ordinary Least Square (OLS) 62 estimates 99 estimation 144 estimator 57, 64, 82, 85–6, 115, 160 orthogonal rotation 34 PANEL.CSV 103 data set 75–6 panel data 3 dataset 324–5 parallel measurements 29 partially restricted residual sum of squares model (PRSS) 68, 70, 72 part-worth coefficients, estimation of 143–4 PC-MDS 143 software 261 perceptions, underlying dimensions of 260 performance information 324 point estimation 57–9 pooling tests and dummy variable models 67–9 strategy for 69–71 Population Discrepancy Function Value (F0) 231, 245 preference data, analysis of 253, 260–1 problem definition 255 PREFMAP 261, 306 example of 292–306 price–quantity map 88 price sensitivity 316 Principal Component Analysis 6, 37–8 component loadings for 42 and Factor Analysis difference between 38 probability density function of error vector 57 probit models 112 PROC SYSLIN 93 product characteristics 315 PROFIT see Property Fitting Property Fitting (PROFIT) 259 analysis 266 example of 275–85 procedures 258 proximity data 306 analysis of 5 proximity, derived measures of 254 proximity matrices 253 alternative 254–5 psychographics 318

Index purchase behavior 320 decision process 320 intentions 165 quadratic effect coding 141 quantal choice models 112–21 fit measures 120 range constraint problem 115 rank order 253 data 6, 137 rank-ordered dependent variables 144 rank-ordered dissimilarities 256 rank-ordered measures of dissimilarity 262 Rao’s R 18, 24 ratio scale 3–4 reflective indicators 38 reliability 29–30, 53 coefficient 44 measurement of 29 of a two-component scale 31–3 restricted residual sum of squares (RRSS) 66 reversed regression 161–2 rotation to terminal solution 42 sales and advertising expenditures, marketing example with 89 for brand as function of price model 76 or market share 104 sample centroids multivariate distribution 13 sampling distribution of 12–13 univariate distribution 12 sample distribution theory 168 SAS 2, 19, 21, 42, 44–5, 47, 49, 52–3, 103, 144, 236 and 2SLS 100 for analysis of variance 52 for computation of means and coefficients 46 examples using 93–103 GLM procedure 143 input to perform the test of a mean vector 21 procedure REG 71 procedure SYSLIN 93, 95–6 in regression analysis 74 for reliability-coefficient 48, 50 for scale construction 51–2 and SUR estimation 94, 97–9 working data set 19

333

“saturated” model 169 scale construction, procedure for 43 scaling factor 256 scales of measurement and their properties 4 types of 3–5 SCAN.DAT 325, 327 scanner data 3, 325 SCANNER.PAS 326–7 statistical independence, model of 169 Seemingly Unrelated Regression (SUR) 79–83 estimator 99 example 93 segmentation 53 Shepard diagram 262 significance test K -sample problem 17–18 one-sample problem 13 two-sample problem 15–17 univariate test 13 similarity data, analysis of 6 in management research 253 problem definition 255 similarity judgments, individual differences in 260 simultaneity and identification 88–91 simultaneous equations, system of 83–8 single equation econometrics 6 social sciences research in 162 Sonites 315–17 space geometry 33 specialty retail stores 316 standard regression model with categorical dependent variables 112 statistical analysis software 42 statistical inference, least squares and maximum likelihood 55–65 statistical tables 310–13 statistics of fit 120 stimuli, minimum number of 257 stress as an index of fit 256 structural model with measurement models, example of 221–50 structural relationship parameters 171 subject preferences and stimuli 306 subsets of data, pooling issues 65–71 sum of squares cross products (SSCP) matrix 14, 16–17 supply and demand curves 88 supply and demand inter-relationships 88 SURVEY.ASC Data File 53 SURVEY data 154, 250

334 survey data 3 questionnaire and scale type 318–21 coding of variables 322–3 symmetric (half) matrix-missing diagonal 254 SYSLIN 101 procedure and 2SLS 100–1 system of equations econometrics 6 Three Stage Least Squares (3SLS) 87–8, 92–3 example 101–3 total SSCP matrices 18 transformational logit 113–17 Two Stage Least Squares (2SLS) 87, 92–3 example 99–103 two-groups discriminant analysis 128 unbiasedness 61 underlying perceptual dimensions 306

Index unidimensionality 53 verification of 29 unique components 40 univariate normal distribution 9 unordered model 147 unrestricted residual sum of squares (URSS) 66 variance maximizing rotations 34–7 VARIMAX rotation method 42 vector and matrix differentiation 309 model of preference 260–1, 293 Wilk lambda of 18 likelihood-ratio criterion of 17 Windows 130 Explorer 148 Wold non-linear iterative least squares procedure of 260

ERRATA The publisher regrets the errors that appear in the first printing of this book. Below please find the corrected text.

Page 60: Equations 4.29 and 4.32: should replace the delta sign with the differentiation operator both at the denominator and numerator (i.e., correct the sign) Page 62: Equations after "BLUE (Best Linear Unbiased Estimator)": should have no small "k" but only capital "K" (2 occurences in equation and 2 other occurences in 2 lines above it) Page 64: Equation 4.54: should have no small "k" but only capital "K" Page 65: Equation 4.63: should have no small "k" but only capital "K" Page 86: Equation 5.43: should end with "y" sub "1" (i.e., the sub "1" is missing) Page 167: Fig. 8.1.: upper left hand corner part of figure: the three numbers "1", "2" and "3" should be preceeded with the delta sign (i.e., this sign is missing); the equation should then read: "delta" sub "1", "delta" sub "2", "delta" sub "3"